About This Week’s Covers

This week’s cover image is inspired by the fun milestone of reaching 50,000 links, manually organized into 59 categories over 129 weeks!

It’s important to remind anyone reading this that I have a complete and full life beyond my newsletter. I simply replace a lot of dilly-dallying with productivity. I touch a lot of grass, sand, snow, and water.

I still fully enjoy my family, the gym, my job, and my hobbies. In no way do I associate my identity with my newsletter, other than I really like doing it, and boy am I learning a lot about operations, automation, agents, and APIs. The best way to learn is through play.

Back to the cover image… Because artificial intelligence is trained on existing material, it’s always recycled. So I had little fun by creating a play on the Audi 5000, and I made an advertisement using a rusty old Audi instead of a brand-new one. I recreated an ad from the ’80s and changed the 5000 to 50,000 and added the word LINKS.

When I was a kid we said “Audi 5 G’s” as a goodbye (shoutout to Big Daddy Kane).

The original Audi image

The rest of the category covers are a test of Gemini’s dynamic image-editing abilities using the API. I gave Gemini a photo of a Mercedes hood ornament and then ran my automation script to have Claude determine what each of my categories could be transformed into as a hood ornament to swap out the Mercedes one.

The original hood ornament image

It did a pretty good job. My favorite few are included below:


This week’s humanities reading is by Hesiod, from the ancient Greek didactic poem Works and Days. It celebrates the accumulation of hard work over time, in the spirit of the 50,000 links.

“He who adds to what he has, will keep off bright-eyed hunger; for if you add only a little to a little and do this often, soon that little will become great.”

-Hesiod, Works and Days

This Week By The Numbers

Total Organized Headlines: 457

This Week’s Executive Summaries

This week, I organized 457 links, 116 of which informed the executive summaries. I’m going to start with top stories and then go through the rest alphabetically by company name.

Top Stories

NVIDIA I

Jensen’s $1,000,000,000,000 Prediction
“Breaking: 1 trillion revenue for NVIDIA in 2027 Jensen Huang: “One year after last GTC, right here where I stand… I see, going down so much, through 2027. At least… one trillion dollars, you know? Now, does it make any sense? I’m certain computer demand will be much https://x.com/TheTuringPost/status/2033622628385362068

“Jensen just said NVIDIA’s $1T projection for 2025-27 covers only Blackwell and Rubin to keep it consistent with the previous projection. He mentioned he could have included Groq in that number: “”””so if I would’ve included that, theoretically, not actually, but theoretically, https://x.com/TheHumanoidHub/status/2033990614824665421

“Nvidia targets data center revenue of $1+ trillion for 2025-2027. That’s already quite ridiculous, with the AI physical world only in its zeroth innings . $NVDA https://x.com/TheHumanoidHub/status/2033627322331660784

Jensen’s $50,000,000,000,000 Prediction
“Jensen is cementing the idea that Nvidia-powered AI is now the backbone of every major industry. He said robotics alone will be a $50 trillion industry. https://x.com/TheHumanoidHub/status/2033619022508659118

Palantir

Palantir/Department of War: Maven AI Walk-through
“Cameron Stanley, Chief Digital and Artificial Intelligence Officer of the Department of War, shares how the DoW is driving enterprise-wide adoption of data, analytics, and AI to generate decision advantage—and what it takes to move cutting-edge technology from the lab to the warfighter at speed.”

I highly recommend watching the full ten minute video:

“Probably the most current look at Palantir’s maven smart system software. Here’s the DoW’s Chief AI officer showing how it works: https://x.com/bilawalsidhu/status/2032432668105712093

“Palantir on X: “”””This is Maven Smart System—Palantir’s software as a service product that we are deploying across the entire department.”” https://t.co/hIaQAiq4iJ / X https://x.com/PalantirTech/status/2032142543022960980

Model Adoption Rate

Ramp AI Index
“Ramp AI Index shows overall business AI adoption rose to 47.6% of businesses in February, a record high, with 24.4% of businesses now using Anthropic. OpenAI’s adoption rate fell by 1.5%.

Anthropic adoption grew 4.9% month over month, its largest monthly gains since we started tracking. Nearly one in four businesses on Ramp now pays for Anthropic (a year ago, it was one in 25). OpenAI’s 1.5% decline was the largest in any single month for any AI model company since we started tracking business AI adoption. Google adoption grew slightly to 4.7% while xAI held at less than 2% of businesses.”
https://ramp.com/velocity/ai-index-march-2026

Molmo

MolmoPoint: Better pointing architecture for vision-language models https://allenai.org/blog/molmopoint

“spent some time today playing with MolmoPoint it’s pretty crazy that we can use VLMs for multi-object tracking now instead of spelling out coordinates as text, it points by directly selecting parts of its own visual features prompt: “”””Track blue players.”””” https://x.com/skalskip92/status/2034606226902827228

Molmo is a little-known yet incredible video understanding model from the Allen Institute for AI. The Allen Institute is a nonprofit scientific research institute that was founded by former Microsoft co-founder Paul Allen 16 years ago. Most of Molmo’s products are completely open source. I first saw Molmo appear in week 52, back in September of 2024. It didn’t really make another executive summary until week 116 in December of 2025. Molmo came out of the gate with this really cool family of open, state-of-the-art multimodal AI models that were incredibly efficient.

Molmo has always had a striking ability to understand videos and pictures in almost uncanny ways, the way that humans do. You can show it a picture of a drawing or a table of data, and it will turn that into a structured JSON file. Even though it’s almost unheard of, it often outperforms OpenAI, Gemini, and Claude.

Traditional tracking and segmentation

But just a few months ago, in December, Molmo came out with an absolutely mind-blowing second edition called Molmo 2. It goes beyond just the idea of object segmentation or tracking. Of course, it can follow and track objects, but it also has the real context of a video.

One of my favorite examples from the Molmo release blog is that you could have Molmo watch a video of someone cooking, and Molmo will actually extract the recipe instructions by watching the video. It will watch the video, observe someone cooking, and along the way understand what to put into the recipe, how much, and when. It can count objects, track videos, and answer complex questions about what’s happening. I encourage you to go back and watch all of the Molmo demonstrations from December.

This week, the latest Molmo model is called Molmo Point.

Claude explains the “Point” part of the release nicely (way better than I could!):

What is “pointing” in AI? Imagine you show an AI a photo and ask “where’s the coffee mug?” A basic AI might say “there’s a mug in the image.” But a grounded AI can actually point to it — like saying “it’s right there,” and meaning a specific pixel location. That’s what this is about. Pointing matters a lot for things like:

Robots figuring out where to grab something AI agents clicking the right button in an app Counting or tracking objects across video frames

The old way (the problem) Most models pointed by generating coordinates as text — essentially spelling out numbers like “x=342, y=178.”

Think of it like giving directions by reciting GPS coordinates out loud instead of just pointing your finger. It works, but it’s clunky:

The AI had to learn an entirely separate coordinate system It used a lot of “words” (tokens) just to say “over there” It got fragile at high resolutions

The new way (MolmoPoint) Instead of spelling out coordinates, MolmoPoint lets the model point by directly selecting parts of its own visual input — the actual image features it’s already looking at. allenai Here’s the analogy: imagine the AI is already holding a mental map of the image in its head. The old way forced it to translate that mental map into written coordinates. The new way lets it just tap the map directly. It does this in three steps — zoom out, zoom in, pinpoint — like how you’d find something on Google Maps: country → city → street.

“It’s like realizing you’ve been giving someone written directions when you could have just handed them a map with an X on it. The information was always there — the delivery method was just unnecessarily complicated.”

NVIDIA II

DLSS 5
NVIDIA’s new DLSS 5 is basically very quick upscaling of imagery. If you’ve ever used a tool like Magnific (https://magnific.ai/editor/), you know the before-and-after look of a low-resolution image that’s been upscaled with artificial intelligence. This is literally the same thing, but with a little more consistency across frames and at real-time speeds for gaming and rendering.

NVIDIA DLSS 5 Delivers AI-Powered Breakthrough In Visual Fidelity For Games https://www.nvidia.com/en-us/geforce/news/dlss5-breakthrough-in-visual-fidelity-for-games/

NVIDIA DLSS 5 Announcing NVIDIA DLSS 5, an AI-powered breakthrough in visual fidelity for games, coming this fall. DLSS 5 infuses pixels with photorealistic lighting and materials, bridging the gap between rendering and reality. Learn More ttps://x.com/NVIDIAGeForce/status/2033617732147810782

“DLSS 5 is completely mind blowing. The neural rendering model with photoreal lighting and materials is a generation step up in visual fidelity. Gaming with DLSS 5 feels like future tech, but its possible now. It is truly incredible. 🤯 https://x.com/GeForce_JacobF/status/2033615891045454112

“DLSS 5 might be the moment where the anti AI pendulum starts swinging back. Many in the 3D community who were against generative AI are now pushing back on the “”””everything is AI slop”””” crowd. The pendulum swung too far and they can feel it. Nice to see the rebalancing. https://x.com/bilawalsidhu/status/2034281398052274666

“Here’s everything we know about Nvidia’s “”””greatest leap in graphics since real-time ray tracing”””” You can see Digital Foundry’s jaw drop in this reaction after they just saw DLSS 5.0: – Will ship in Fall of 2026! – Demo ran 4k on 2 5090’s but is already running on single GPU in https://x.com/Grummz/status/2033641075806769382

Niantic

Companies are building maps of the world
Bilawal Sidhu did a great job summarizing the various ways companies are tracking and building maps of the world:

“People are undoubtedly a little alarmed at having unwittingly helped build a 3D map of the world for Niantic by contributing 30 billion crowdsourced images.”

“And lest we forget that Niantic is just one of many companies quietly building their own map of the world right now — and they’re all capturing different facets of reality:”

>🚶 person-level: Axon body cams on hundreds of thousands of officers. Meta Ray-Ban glasses capturing first-person POV at scale — overseas operators reviewing images every time someone says “Hey Meta.”

> 🚗 vehicle-level: Tesla dashcams on every car in the fleet, massive onboard compute extracting and distilling data to the cloud. Waymo with cm-accurate 3D maps of every city they operate in. Fleet telematics cameras on delivery vehicles globally.

> 🏠 street & home-level: Flock Safety deploying CCTV across neighborhoods and cities. Amazon with Ring cameras on every doorstep and mailroom (recently got dragged over that Super Bowl commercial about fusing all these cams together to find your dog) plus dashcams on every Prime delivery van. Roomba mapping your floor plan every time it vacuums — Amazon wanted that data badly enough to try acquiring iRobot for $1.7B before regulators shut it down.

> 🥽 headset-level: Apple Vision Pro and Meta Quest build a 3D model of whatever room you’re in every time you put them on. Between Ring, Roomba, and your headset, your entire home is being spatially understood by at least three different companies.

>📍platform-level: Google with Street View cars, aerial planes, satellite imagery, and live location from every Android phone in your pocket. Apple doing the same with mapping cars AND every LiDAR iPhone is quietly a 3D scanner. And yeah, despite the “Apple is too privacy-conscious” narrative, they’re collecting location data too.

>🏃 trajectory-level: Strava mapped every running and cycling trail on Earth — and accidentally exposed secret military bases in Afghanistan and Syria because soldiers logged their jogs. When you aggregate enough individual trajectories, patterns emerge that were never supposed to be visible.

> 🛰️ space-level: Planet Labs imaging the entire Earth’s landmass every single day from orbit. Vantor capturing it in higher detail. Iceye doing it in 3D using SAR. If something changes anywhere on the planet — a building goes up, a forest burns down, a military convoy moves — before-and-after imagery within 24 hours.

Fused together — we have everything from body cam to dashcam to doorbell to phone to satellite — every layer of physical reality is being mapped by somebody right now. Different sensors, different angles, different purposes. Same pattern. “
https://x.com/bilawalsidhu/status/2033350363982471182

Google

Measuring progress toward AGI: A cognitive framework
We’re introducing a framework to measure progress toward AGI, and launching a Kaggle hackathon to build the relevant evaluations. https://blog.google/innovation-and-ai/models-and-research/google-deepmind/measuring-agi-cognitive-framework/

Deconstructing general intelligence
Our framework draws on decades of research from psychology, neuroscience and cognitive science to develop a cognitive taxonomy.

It identifies 10 key cognitive abilities that we hypothesize will be important for general intelligence in AI systems:

Perception: extracting and processing sensory information from the environment
Generation: producing outputs such as text, speech and actions Attention: focusing cognitive resources on what matters
Attention: focusing cognitive resources on what matters
Learning: acquiring new knowledge through experience and instruction
Memory: storing and retrieving information over time
Reasoning: drawing valid conclusions through logical inference
Metacognition: knowledge and monitoring of one’s own cognitive processes
Executive functions: planning, inhibition and cognitive flexibility
Problem solving: finding effective solutions to domain-specific problems
Social cognition: processing and interpreting social information and responding appropriately in social situations

OpenAI

Private Equity
“This news came out a little earlier than we planned; we’re excited to be building a deployment arm and will share more details soon.

Companies have a ton of urgency to deploy AI in their organizations and we’re sprinting to meet that demand. More than 1 million businesses run on OpenAI products. Codex is now at 2M+ weekly active users, up nearly 4x since the start of the year. API usage jumped 20% in the week after GPT-5.4 launched. And Frontier, which launched last month to help enterprises build, deploy, and manage AI coworkers that can do real work, has way more demand than we can handle.

That’s why we launched Frontier Alliances so we leverage our ecosystem of partners to scale. And that is also why we are launching a dedicated deployment arm tasked with embedding Forward Deployed Engineers deeply inside of enterprises. This project has been in the works with our investor and alliance partners since last December, and we are grateful for them and their partnership.”
https://x.com/fidjissimo/status/2033537381907710092

Rescue Dog Tumor
“An AI consultant with no biology training used ChatGPT and AlphaFold to create a personalized mRNA cancer vaccine for his rescue dog. Tumor shrunk by half. UNSW structural biologist Dr. Kate Michie: “It’s exciting to me that someone who’s not a scientist has been able to do https://x.com/TheRundownAI/status/2032843584869708105

“this is actually insane > be tech guy in australia > adopt cancer riddled rescue dog, months to live > not_going_to_give_you_up.mp4 > pay $3,000 to sequence her tumor DNA > feed it to ChatGPT and AlphaFold > zero background in biology > identify mutated proteins, match them to https://x.com/IterIntellectus/status/2032858964858228817

“How AI empowered Paul Conyngham to create a custom mRNA vaccine to cure his dog’s cancer when she had only months to live. The first personalized cancer vaccine designed for a dog: https://x.com/gdb/status/2032867435704103006

Selling A House
“Florida man sold his house in just 5 days after letting ChatGPT handle the entire process instead of a real estate agent The AI handled pricing, marketing, showings, and even helped draft the contract https://x.com/i/birdwatch/t/2032864183918690675?source=6

“i mean this story is insane. man used chatgpt to sell his house in 5 DAYS. got 5 offers in 72 hours. no real estate agents. saved so much money doing it too. he used AI to: > price the house (researched neighboring properties for sale) > wrote up the legal contracts (saving https://x.com/cryptopunk7213/status/2033194801852567620?s=46

Man uses ChatGPT to sell his Cooper City home – NBC 6 South Florida https://www.nbcmiami.com/news/local/innovation-on-6/man-uses-chatgpt-to-sell-his-cooper-city-home-it-exceeded-our-expectations/3778919/

Microsoft

Co-Pilot Health
Introducing Copilot Health | Microsoft AI https://microsoft.ai/news/introducing-copilot-health/

“We’re approaching the dawn of medical superintelligence – the moment when affordable, world-class medical knowledge and support is at your fingertips whenever you need it. I think people are still underestimating how profound this transformation is going to be. Today we’re https://x.com/mustafasuleyman/status/2032092644483141928

“I’m not saying Copilot diagnosed me. I’m saying it helped me ask for the right test. A test no doctor had ordered for me in twenty years.”””” This is why I’m so passionate about AI & healthcare. https://x.com/mustafasuleyman/status/2033655842919395723

Cancer Lab Work
“Every time you get a cancer biopsy, the lab makes a tissue slide that costs about $5. It shows the shape of your cells under a microscope, and every cancer patient already has one on file. There’s a much fancier version of that test called multiplex immunofluorescence (basically https://x.com/anishmoonka/status/2033344818475360562

GigaTIME: Scaling tumor microenvironment modeling using virtual population generated by multimodal AI – Microsoft Research https://www.microsoft.com/en-us/research/blog/gigatime-scaling-tumor-microenvironment-modeling-using-virtual-population-generated-by-multimodal-ai/

Anthropic

Channels
“Use channels to push messages, alerts, and webhooks into your Claude Code session from an MCP server. Forward CI results, chat messages, and monitoring events so Claude can react while you’re away.

A channel is an MCP server that pushes events into your running Claude Code session, so Claude can react to things that happen while you’re not at the terminal. Channels can be two-way: Claude reads the event and replies back through the same channel, like a chat bridge. Events only arrive while the session is open, so for an always-on setup you run Claude in a background process or persistent terminal.”

“Push events into a running session with channels – Claude Code Docs https://code.claude.com/docs/en/channels

“Read about the feature and how to build your own channels for Claude Code on our docs! https://x.com/neilhtennek/status/2034762489951658190

“Today we’re launching channels for Claude Code as an experimental feature! A few days ago, I was fed up that I couldn’t text Claude on the go like I would any of my friends. But those days are gone! Claude is saved in my contacts and I can keep shipping on the go. https://x.com/neilhtennek/status/2034762196576805123

Cowork Dispatch (More to follow next week!)
“We’re shipping a new feature in Claude Cowork as a research preview that I’m excited about: Dispatch! One persistent conversation with Claude that runs on your computer. Message it from your phone. Come back to finished work. To try it out, download Claude Desktop, then pair https://x.com/felixrieseberg/status/2034005731457044577

“OpenClaw, Anthropic version. Basically: assign work from anywhere, come back to finished results. This is what “”””AI assistant”””” was always supposed to mean. https://x.com/fdaudens/status/2034080669119152238

“We’re shipping a new feature in Claude Cowork as a research preview that I’m excited about: Dispatch! One persistent conversation with Claude that runs on your computer. Message it from your phone. Come back to finished work. To try it out, download Claude Desktop, then pair https://x.com/felixrieseberg/status/2034005731457044577?s=12

“After using it a bit, Claude Cowork Dispatch covers 90% of what I was trying to use OpenClaw for, but feels far less likely to upload my entire drive to a malware site. https://x.com/emollick/status/2034067677157679379

A Guide To Using Skills
Lessons from Building Claude Code: How We Use Skills https://x.com/trq212/status/2033949937936085378

NVIDIA III

LangChain 1B Downloads
“Announced in Jensen’s keynote today: LangChain frameworks have crossed 1B downloads. We’re excited to join the NVIDIA Nemotron Coalition to help shape the open models that power these agents. ➡️ Read the announcement: https://t.co/CWlbAzhlXy ➡️ Check out the docs: https://x.com/LangChain/status/2033788913937195132

OpenAI II

Subagents
“A knowledge-work platform built around GPT-5.4 Pro level intelligence would be really useful. The gap between other models and what Pro can do on complex intellectual work remains stark. I would love to have access in a Codex-like platform with shared file spaces, subagents, etc https://x.com/emollick/status/2033959257196966360

“GPT-5.4 mini matters for subagents because it changes what feels worth handing off. The parent thread should hold the architecture, plan, and progress narrative. Fast subagents can explore the repo, check hypotheses, and preserve the parent thread’s limited attention. https://x.com/nickbaumann_/status/2034134875234832540#m

“We’re introducing GPT-5.4 mini and nano, our most capable small models yet. GPT-5.4 mini is more than 2x faster than GPT-5 mini. Optimized for coding, computer use, multimodal understanding, and subagents. For lighter-weight tasks, GPT-5.4 nano is our smallest and cheapest https://x.com/OpenAIDevs/status/2033953815834333608

“Subagents are now supported in Codex. They’re very fun and make it possible to get large amounts of work done *quickly*: https://x.com/gdb/status/2033757784437895367

Customer Service v. AI

The internet ruined customer service. AI could save it. | Andreessen Horowitz
https://a16z.com/the-internet-ruined-customer-service-ai-could-save-it/

“You, reader, have been on hold for 45 minutes waiting to talk to a human who can solve your problem. You’ve been transferred to another department. You’ve screamed “representative” into an automated phone tree. You’ve rage-typed into an unintelligent chatbot that isn’t programmed to know what you’re angry about.

And here’s the strange thing—as businesses have gotten better at everything else, customer experience has gotten worse. That’s one of the great ironies of the internet age. The internet helped commerce scale infinitely: one-click purchasing, next-day delivery, oceans of choices for practically any consumer good. But it made the experience of being a customer a lot worse. The more customers a business services, the less attention any individual customer gets: because while logistics can scale exponentially, devoted customer attention—real attention, to make sure you’re happy with the experience—can only scale linearly. That is to say, with the number of humans that the business employs.

For any business that sells goods online, the inevitable fact is that customer service is a cost center to be minimized, not a relationship to be deepened. Each customer becomes a ticket number, a case ID, a position in an endless queue. This isn’t because businesses today care less than businesses did 100 years ago or 4,000 years ago. It’s simply because devoted attention doesn’t scale.

Hermès can afford to give you a concierge, because you’re one of a relatively small number of people who are able to spend five or six figures a year there, and because the “personal touch” is at the core of what Hermès provides. Delta and Verizon can’t make the same call. So the rich get the concierge; normal people get the phone tree. That’s the basic deal for the consumers of scale businesses: you get the scale pricing, but you lose the intimacy. And that essential tradeoff has governed consumer business for decades.

But AI collapses the cost of high-quality attention. When you have hyperintelligent AI that is ultra-cheap, always on, and can run infinite parallel instances—each with a full memory of the customer’s history, preferences, and context—then the entire logic of customer engagement flips in an instant. The question is no longer “can we afford to pay attention to this customer?” It’s “what becomes possible when the marginal cost of a customer interaction approaches zero?”

The answer, I believe, is that every business becomes a concierge business. When attention becomes abundant, customer experience undergoes a fundamental transformation.”

The Rest of the Week’s AI News

Anthropic

DoW
Opposition/Response to Motion – #96 in Anthropic PBC v. U.S. Department of War (N.D. Cal., 3:26-cv-01996) – CourtListener.com https://www.courtlistener.com/docket/72379655/96/anthropic-pbc-v-us-department-of-war/

Usage Survey
“We invited Claude users to share how they use AI, what they dream it could make possible, and what they fear it might do. Nearly 81,000 people responded in one week—the largest qualitative study of its kind. Read more: https://x.com/AnthropicAI/status/2034302152945144166#m

What 81,000 people want from AI \ Anthropic https://www.anthropic.com/features/81k-interviews

Bezos

Exclusive | Jeff Bezos in Talks to Raise $100 Billion for AI Manufacturing Fund – WSJ
https://www.wsj.com/tech/jeff-bezos-aims-to-raise-100-billion-to-buy-revamp-manufacturing-firms-with-ai-618a3cfe

Business

OCR
“NEW SOTA OCR MODEL DROPPED Congrats to @VikParuchuri and team for releasing Chandra OCR 2! – 85.9% on olmocr bench, making it first place 🏆 – 90+ language support – 4B model – Full layout information – Extracts + captions images and diagrams – Strong handwriting, math, form, https://x.com/nathanhabib1011/status/2034565076963991910

ByteDance

SeeDance Paused!
Just a few weeks ago, I published a newsletter, and the cover image was a video created by Chinese director Jia Zhangke. Zhangke used ByteDance’s Seedance 2.0 to make a five-minute video that’s almost impossible to tell is AI. He even purposely made a few bits look like AI because it plays into the movie.

Seedance was considered the first video tool that successfully crossed the uncanny valley to the point where it was unrecognizable as artificial intelligence. Deepfakes started to proliferate, and folks who track AI started to really voice concerns that we were going to be in trouble because nobody can tell the difference between real and fake on Seedance. That’s why this week’s news is such a big deal, because ByteDance had to put the kibosh on Seedance 2.0 after a series of lawsuits from Hollywood over copyright violations.

ByteDance reportedly pauses global launch of its Seedance 2.0 video generator | TechCrunch
https://techcrunch.com/2026/03/15/bytedance-reportedly-pauses-global-launch-of-its-seedance-2-0-video-generator/

ByteDance Suspends Launch of Video AI Model After Copyright Disputes With Hollywood — The Information https://www.theinformation.com/articles/bytedance-suspends-launch-video-ai-model-copyright-disputes-hollywood

Protein Folding
Here is SeedProteo, our latest diffusion-based model for de novo all-atom protein design from ByteDance Seed! Our server is now live — feel free to give it a try!
https://x.com/SeedFold/status/2033515503839514771

Data-Driven Interfaces

US Job Market Visualizer
https://karpathy.ai/jobs/

Andrej Karpathy is a pioneering AI researcher who coined the term “vibe coding.” He built a U.S. job market visualizer that uses the Bureau of Labor Statistics’ Occupational Outlook Handbook and leverages large language models to scrape and parse occupations by a variety of criteria. It’s an incredibly dynamic website that shows just how powerful vibe coding can be when brought up against public data sources. I highly recommend clicking around and looking especially at the Digital AI Exposure tab.

DoorDash

DoorDash’s New Paid Tasks Turn Couriers Into AI and Robot Trainers – Bloomberg
https://www.bloomberg.com/news/articles/2026-03-19/doordash-s-new-paid-tasks-turn-couriers-into-ai-and-robot-trainers

Education

Humanities
“Never been a better time to study the humanities:
1) LLMs are trained on the cultural history of all humans, knowing that helps you use them
2) The humanities gives us context in this odd moment in history
3) Books & stuff are good Wrote this 3 years ago: https://x.com/emollick/status/2033692979844485304

Excel

Agent Excel Bakeoff
“Hey Excel agents from Claude, OpenAI & MS Copilot: “”””make me a working strategy game in excel, it should have some form of graphics”””” Claude made a board and acted as game master, Copilot created a board but no game, ChatGPT built a working game with formulas with a “”””smart”””” enemy.
https://x.com/emollick/status/2033372471395512566

Google

Design
Design UI using AI with Stitch from Google Labs https://blog.google/innovation-and-ai/models-and-research/google-labs/stitch-ai-ui-design/

Exclusive: Early look at upcoming design tool from Google https://www.testingcatalog.com/exclusive-early-look-at-upcoming-vibe-design-tool-from-google/

Personal Intelligence
Personal Intelligence in AI Mode and Gemini expands in the U.S. https://blog.google/products-and-platforms/products/search/personal-intelligence-expansion/

Vibe Coding
“Introducing the all new vibe coding experience in @GoogleAIStudio, feating: – One click database support – Sign in with Google support – A new coding agent powered by Antigravity – Multiplayer + backend app support and so much more coming soon! https://x.com/OfficialLoganK/status/2034656376450908203

Vibe Code to production with Google AI Studio https://blog.google/innovation-and-ai/technology/developers-tools/full-stack-vibe-coding-google-ai-studio/

Karpathy’s AutoResearch Project

A Guide to Andrej Karpathy’s AutoResearch: Automating ML with AI Agents | DataCamp
https://www.datacamp.com/tutorial/guide-to-autoresearch

“Learn how Karpathy’s AutoResearch runs 100+ ML experiments overnight on a single GPU. Covers the three-file architecture, ratchet loop, results, and limitations. AutoResearch is an open-source Python tool that lets an AI agent run ML experiments on a single GPU without human intervention. It loops through propose-train-evaluate cycles, keeping only changes that improve validation loss. The project ships under an MIT license.”

“Karpathy’s Autoresearch is bottlenecked by a single GPU. We removed the bottleneck. We gave the agent access to our K8s cluster with H100s and H200s and let it provision its own GPUs. Over 8 hours: • ~910 experiments instead of ~96 sequentially • Discovered that scaling model https://x.com/skypilot_org/status/2034681533051855173

“karpathy/autoresearch: AI agents running research on single-GPU nanochat training automatically https://github.com/karpathy/autoresearch

LangChain

Deep Agents
“LangChain just open-sourced Deep Agents—an agent harness that’s opinionated and ready-to-run out of the box. Instead of wiring up prompts, tools, and context management yourself, you get a working agent immediately and customize what you need. It’s an MIT-licensed system that’s https://x.com/itsafiz/status/2033591253955449289

Fleet
“Introducing LangSmith Fleet: an enterprise workspace for creating, using, and managing your fleet of agents. Fleet agents have their own memory, access to a collection of tools and skills, and can be exposed through the communication channels your team uses every day. Fleet https://x.com/LangChain/status/2034679590250258855

“Introducing LangSmith Fleet. Agents for every team. → Build agents with natural language → Share and control who can edit, run, or clone each agent → Manage authentication with agent identity → Approve actions with human-in-the-loop → Track and audit actions with tracing in https://x.com/LangChain/status/2034694530478612777

MCP

Roller Coaster Ride MCP is Dead; Long Live MCP! https://chrlschn.dev/blog/2026/03/mcp-is-dead-long-live-mcp/

“good concepts here in fastmcp on distributing skills via MCP resources, i think this might be the right approach solves the problem of skills going out of date if you load them fresh each time + can tie them to tools easier more to explore here https://x.com/RhysSullivan/status/2034125767987368242#m

“MCP was a mistake. Long live CLIs. https://x.com/skirano/status/2034269154404868314#m

Meta

Layoffs
Exclusive: Meta planning sweeping layoffs as AI costs mount | Reuters
https://www.reuters.com/business/world-at-work/meta-planning-sweeping-layoffs-ai-costs-mount-2026-03-14/

MidJourney

V8
“Today we’re starting to test an early version of our V8 model with our community. It’s much better at following prompts, 5x faster, has native 2K modes, improved text rendering and the best personalization, sref, and moodboard performance ever. Have fun! https://x.com/midjourney/status/2034015403542974793?s=20

MiniMax

Self-evolving models
MiniMax M2.7: Early Echoes of Self-Evolution – MiniMax News | MiniMax https://www.minimax.io/news/minimax-m27-en

“BREAKING 🚨: MiniMax released MiniMax M2.7, a new self-evolving model, achieving a score of 56.22% on SWE-Bench Pro. M2.7 was used for building complex agent harnesses during its own development. Users can now access MiniMax M2.7 via APIs and MiniMax Agent. https://x.com/testingcatalog/status/2034250919345377604#m

“During the iteration process, we also realized that the model’s ability to recursively evolve its harness is equally critical. Our internal harness autonomously collects feedback, builds evaluation sets for internal tasks, and based on this continuously iterates on its own https://x.com/MiniMax_AI/status/2034315323109953605#m

“Introducing MiniMax-M2.7, our first model which deeply participated in its own evolution, with an 88% win-rate vs M2.5 – Production-Ready SWE: With SOTA performance in SWE-Pro (56.22%) and Terminal Bench 2 (57.0%), M2.7 reduced intervention-to-recovery time for online incidents https://x.com/MiniMax_AI/status/2034315320337522881#m

MiniMax Global Announces Full Year 2025 Financial Results – MiniMax News | MiniMax https://www.minimax.io/news/minimax-global-announces-full-year-2025-financial-results

“Minimax M2.7 released! And its a big one Highlights: Self-evolving – first model that helped build itself, running 100+ autonomous optimization loops during its own RL training (30% internal improvement). Strong coder – 56.2% on SWE-Pro (near Opus 4.6), 55.6% on VIBE-Pro, https://x.com/kimmonismus/status/2034269026353082422#m

Mistral

Forge
Introducing Forge | Mistral AI
https://mistral.ai/news/forge

Leanstral
Leanstral: Open-Source foundation for trustworthy vibe-coding | Mistral AI https://mistral.ai/news/leanstral

Small
Introducing Mistral Small 4 | Mistral AI
https://mistral.ai/news/mistral-small-4

Moonshot

Attention Technical Paper
This is a complicated topic, so I am providing (vetted and approved) xplanations from GPT and Claude:

GPT:

The old way

A Transformer has many layers, maybe 40, 80, or more.
Each layer looks at the current representation of the text, does some thinking, and then adds its result back on top.

So it is kind of like this:

  • Layer 1 adds its note
  • Layer 2 adds its note
  • Layer 3 adds its note
  • Layer 4 adds its note

and so on.

That sounds reasonable, but there is a catch:

every old note stays in the pile with equal weight.

So by the time you get very deep into the model, the current state is like a giant stack of all prior layers’ contributions.

One-sentence summary

In normal Transformers, every layer keeps dumping its output into one ever-growing bucket, which can drown out later useful work. Kimi changed that so the model can selectively retrieve the most relevant earlier layer information instead of treating all past layers equally.
https://chatgpt.com/c/69d596c6-0e28-8325-a1ef-43e00697d4a1

Claude:

The practical problem and the fix
Full attention across all 40+ layers is expensive — you’d have to keep every floor’s output in memory simultaneously during training across many GPUs. That’s a lot of data moving around.

So they introduced Block Attention Residuals: group the layers into ~8 chunks (blocks). Within each chunk, use the old simple pile-stacking. But between chunks, use the smart attention mechanism to decide how much each chunk’s summary matters.

It’s like summarizing chapters of a book. Within a chapter, you read everything sequentially. But when writing your final essay, you selectively draw on whichever chapter summaries are actually useful — not all of them equally.

One bonus: the original raw input (the actual words, before any processing) is always kept as a separately accessible source. So any layer, anywhere in the network, can reach all the way back to “what did the user actually say?” — which is surprisingly powerful.

The results, in plain English

It performs as well as a model trained on 25% more data, for free
It adds less than 2% to inference time (basically unnoticeable in production)
Across every benchmark they tested — math, coding, reasoning, general knowledge — it improved
https://claude.ai/share/7239e73e-9e9d-469a-bbdb-e5c7da75a4e9

“🔥 @Kimi_Moonshot’s new Attention Residual paper is sparking discussions. Zhihu contributor OpenLLMAI shares a deep dive: “”””From Kimi’s Attention Residual to ‘Vertical Attention’ — an idea I’ve been thinking about for half a year.”””” Some interesting thoughts on attention mechanisms https://x.com/ZhihuFrontier/status/2033751367198949865

“Avi Chawla on X: “”Big release from Kimi! They just released a new way to handle residual connections in Transformers. In a standard Transformer, every sub-layer (attention or MLP) computes an output and adds it back to the input via a residual connection. If you consider this across 40+ layers, https://t.co/5i5AN9tzIm / X https://x.com/_avichawla/status/2033472650836914495

“Introducing 𝑨𝒕𝒕𝒆𝒏𝒕𝒊𝒐𝒏 𝑹𝒆𝒔𝒊𝒅𝒖𝒂𝒍𝒔: Rethinking depth-wise aggregation. Residual connections have long relied on fixed, uniform accumulation. Inspired by the duality of time and depth, we introduce Attention Residuals, replacing standard depth-wise recurrence with https://x.com/Kimi_Moonshot/status/2033378587878072424

“visual summary of attention residuals by kimi, beautiful paper https://x.com/eliebakouch/status/2033488233854620007”

“@_avichawla Impressive work from Kimi https://x.com/elonmusk/status/2033528245464047805

Nous Research

Hermes Agents
“Built PrediHermes ✨ a Hermes Agent skill + companion WorldOSINT/MiroFish forks for geopolitical prediction. @NousResearch It pulls 54+ OSINT modules, uses Polymarket to find contracts with clear resolution criteria, then runs MiroFish multi-agent sims to model individual https://x.com/WeXBT/status/2033391568426598608

“Hermes Agent v0.3.0 ☤ 248 PRs. 15 contributors. 5 days. • Real-time streaming across CLI and all platforms • First-class plugin architecture, package and share tools+commands+skills • /browser connect to live Chrome via CDP • @vercel AI Gateway model provider • https://x.com/NousResearch/status/2033877040399831478

“Browser Use is now an official provider for the browser tool in Hermes-Agent – Update to try it out 😉 Use `hermes tools` to set the browser backend. (Note: this requires an API key with them) https://x.com/Teknium/status/2033811117521408078

“Did a small local anime server tool powered by Hermes Agent (@NousResearch). You can: – fully sync your anime list – download torrents from different sources – add tracking & scheduled downloads – auto-manage disk usage – serve to any device within your local wifi and more! https://x.com/rodmarkun/status/2033307437088850102

“I have consecutively spent millions of tokens today with Hermes without breaking anything, where Openclaw would’ve needed several nudges. Both have their merits, but for prod, Hermes nailed it. Amazing job @Teknium https://x.com/populartourist/status/2034653545287348266

“i’ve been using @NousResearch Hermes Agent for about a week and my initial thoughts are it just works out of the box I find it’s memory and learning to be far superior to OpenClaw without augmenting it with QMD or any additional memory systems it’s early and my OC’s handle A https://x.com/austin_hurwitz/status/2033552632241857002

“Told my agent to create a fresh OpenClaw agent in digital ocean according to the instructions. Ran into like 6 issues with the flow from the official docs. Did the same for the @NousResearch Hermes agent – Claude one-shotted it 🙂 https://x.com/0xMasonH/status/2033608276286243323

“trying Hermes agent. setup was significantly easier than openclaw. I can see its already basically a leaner and more opinionated openclaw https://x.com/fuckyourputs/status/2033503910376431728

NVIDIA

HealthCare Robots
The First Healthcare Robotics Dataset and Foundational Physical AI Models for Healthcare Robotics https://huggingface.co/blog/nvidia/physical-ai-for-healthcare-robotics

Karpathy
“🙌 Andrej Karpathy’s lab has received the first DGX Station GB300 — a Dell Pro Max with GB300. 💚 We can’t wait to see what you’ll create @karpathy! 🔗 https://t.co/8ct5QZ3frS @DellTech https://x.com/NVIDIAAIDev/status/2034291235041554871

Jensen Keynote
“What is computer in the future? According to Jensen: “In the future, the computer is really a manufacturing system for tokens. And the number of computers in the world built for token manufacturing is still very small. It’s small because most of the systems we have shipped so https://x.com/TheTuringPost/status/2033983885131059636

“Jensen: “Nvidia is the first vertically integrated but horizontally open company.” This strategy positions Nvidia as the backbone of robotics without stifling innovation. Vertical integration ensures cutting-edge performance on each layer of the AI stack. Horizontal openness https://x.com/TheHumanoidHub/status/2033622691408974133

“GR00T is moving away from VLM-based backbones in favor of integrated world models. Jensen Huang teased GR00T N2 during his keynote; NVIDIA’s next-gen foundation model built on DreamZero research. Utilizing a new world-action model architecture, it succeeds at novel tasks in https://x.com/TheHumanoidHub/status/2034279221372321940

NVIDIA GTC 2026 Keynote: Everything That Happened in 12 Minutes – YouTube https://www.youtube.com/watch?v=X2i_8O75_Os

Video
“A breakthrough in real-time video generation. As a research preview developed with @NVIDIA and shared at @NVIDIAGTC this week, we trained a new real-time video model running on Vera Rubin. HD videos generate instantly, with time-to-first-frame under 100ms. Unlocking an entirely https://x.com/runwayml/status/2034284298769985914#m

World Model
“What if a robot could simulate the physical world from a single image. [📍Bookmark Paper & GitHub for later] PointWorld-1B from Stanford and NVIDIA is a large 3D world model that predicts how an entire scene will move, given RGB-D input and robot actions. The key idea is https://x.com/IlirAliu_/status/2032895393407660380

OpenAI

5.4 mini
Introducing GPT-5.4 mini and nano | OpenAI https://openai.com/index/introducing-gpt-5-4-mini-and-nano/

“GPT-5.4 mini approaches the performance of the larger GPT-5.4 model on several evaluations, including SWE-Bench Pro and OSWorld-Verified. https://x.com/OpenAIDevs/status/2033953828387885470

“GPT-5.4 mini is available today in the API, Codex, and ChatGPT. In the API, it has a 400k context window. In Codex, it uses only 30% of the GPT-5.4 quota, letting you handle simpler coding tasks for about one-third of the cost. GPT-5.4 nano is only available in the API. https://x.com/OpenAIDevs/status/2033953840312291603

“GPT-5.4-mini 2.25 times more expensive than GPT-5-mini $0.75 Input $4.5 Output 400k https://x.com/scaling01/status/2033955279079907511

Astral
OpenAI to acquire Astral | OpenAI https://openai.com/index/openai-to-acquire-astral/

AWS
OpenAI expands government footprint with AWS deal, report says | TechCrunch https://techcrunch.com/2026/03/17/openai-expands-government-footprint-with-aws-deal/

Dictionary
The dictionary sues OpenAI | TechCrunch https://techcrunch.com/2026/03/16/merriam-webster-openai-encyclopedia-brittanica-lawsuit/

IPO
OpenAI preps for IPO in 2026, says ChatGPT must be ‘productivity tool’ https://www.cnbc.com/2026/03/17/openai-preps-for-ipo-in-2026-says-chatgpt-must-be-productivity-tool.html

Shopping
“$WMT is disappointed in results from OpenAI partnership, whereby Walmart users are allowed to shop via ChatGPT and OpenAI would receive a commission on these purchases “Conversion rates—the percentage of users following through with a purchase of an item shown to them by https://x.com/negligible_cap/status/2034369496543305971?s=46

Tutor
“AI really can help education: Randomized controlled experiment on high school students found a GPT-4o powered tutor that personalized problems for students raised final test scores by .15 SD, “”””equivalent to as much as six to nine months of additional schooling by some estimates”””” https://x.com/emollick/status/2033773791688433708

OpenClaw

Chrome browser control
“New @openclaw beta is up: it comes with the new live browser control that Google added in latest Chrome! enable via chrome://inspect#remote-debugging Your clanker will know when to use what, or you can ast it. new “”””user”””” profile session is there! https://x.com/steipete/status/2032686376932491363

Perplexity

Computer
“Computer can now take full control of Comet to complete tasks. When you’re in Comet, Computer spins up a browser agent that can access any site or logged‑in app with your permission, without the need for connectors or MCPs. Available to all Computer users on Comet. https://x.com/perplexity_ai/status/2033598416962592813

Health
Introducing Perplexity Health https://www.perplexity.ai/hub/blog/introducing-perplexity-health

Perplexity launches Perplexity Health agent in US https://www.testingcatalog.com/perplexity-launches-perplexity-health-agent-in-us/

Suno

Fraud
“Someone used Suno AI to generate a Japanese metal band called Neon Oni. Fake member bios, AI-generated music videos, “”””Based in Tokyo”””” on Spotify. 80,000+ monthly listeners. Fans had it in their Spotify Wrapped top 5. Merch was selling. Then, community sleuths exposed it. Traced https://x.com/TheRundownAI/status/2033568236227244451?s=20

Videos

News Tool
Schibsted open sources AI tool that turns news articles into videos | Schibsted https://schibsted.com/news/schibsted-open-sources-ai-tool-that-turns-news-articles-into-videos/

Full Executive Summaries with Links, Generated by Sonnet

Nvidia CEO projects trillion-dollar data center revenue by 2027
Jensen Huang announced Nvidia expects over $1 trillion in data center revenue spanning 2025-2027, driven by demand for AI chips including their Blackwell and Rubin processors. This projection represents unprecedented growth in the semiconductor industry, reflecting the massive infrastructure buildout required for AI deployment. Huang emphasized this conservative estimate excludes newer chip lines, suggesting even higher potential revenue as AI adoption accelerates across industries.

Breaking: 1 trillion revenue for NVIDIA in 2027 Jensen Huang: “One year after last GTC, right here where I stand… I see, going down so much, through 2027. At least… one trillion dollars, you know? Now, does it make any sense? I’m certain computer demand will be much https://x.com/TheTuringPost/status/2033622628385362068

Jensen just said NVIDIA’s $1T projection for 2025-27 covers only Blackwell and Rubin to keep it consistent with the previous projection. He mentioned he could have included Groq in that number: “”so if I would’ve included that, theoretically, not actually, but theoretically, https://x.com/TheHumanoidHub/status/2033990614824665421

Nvidia targets data center revenue of $1+ trillion for 2025-2027. That’s already quite ridiculous, with the AI physical world only in its zeroth innings . $NVDA https://x.com/TheHumanoidHub/status/2033627322331660784

Nvidia CEO predicts robotics will become $50 trillion industry
Jensen Huang is positioning Nvidia’s AI chips as essential infrastructure across all major sectors, with robotics representing the largest potential market opportunity. This signals Nvidia’s strategy to expand beyond current AI applications into physical automation, though the $50 trillion figure far exceeds current global GDP and reflects long-term ambitions rather than near-term projections.

Jensen is cementing the idea that Nvidia-powered AI is now the backbone of every major industry. He said robotics alone will be a $50 trillion industry. https://x.com/TheHumanoidHub/status/2033619022508659118

Anthropic captures 24% of business AI market, overtaking OpenAI momentum
Anthropic now wins 70% of head-to-head matchups against OpenAI among first-time business AI buyers, marking a complete reversal from 2025 trends when OpenAI dominated adoption. Despite charging more for comparable performance and having supply constraints, Anthropic’s cultural positioning as the “cool” alternative to OpenAI’s defense partnerships appears to be driving enterprise preference beyond pure technical metrics. This suggests AI model selection may increasingly resemble consumer brand choices rather than traditional enterprise procurement decisions.

Ramp AI Index March 2026 update https://ramp.com/velocity/ai-index-march-2026

AI2’s MolmoPoint lets vision models point by selecting visual features instead of spelling out coordinates
This breakthrough replaces the clunky method of having AI models generate text coordinates with a more intuitive system that directly selects parts of the model’s own visual understanding. MolmoPoint achieves state-of-the-art results on pointing benchmarks while using 60% fewer tokens per point, and proves especially valuable for robotics, computer interfaces, and video tracking where precise location matters. The approach works because it aligns with how the model actually “sees” rather than forcing it to learn an artificial coordinate system.

MolmoPoint: Better pointing architecture for vision-language models | Ai2 https://allenai.org/blog/molmopoint

spent some time today playing with MolmoPoint it’s pretty crazy that we can use VLMs for multi-object tracking now instead of spelling out coordinates as text, it points by directly selecting parts of its own visual features prompt: “”Track blue players.”” https://x.com/skalskip92/status/2034606226902827228

NVIDIA launches DLSS 5 to make video games look photorealistic
NVIDIA’s new AI technology promises to dramatically improve video game graphics by using artificial intelligence to enhance lighting and materials in real-time. This represents a significant leap beyond previous versions that mainly boosted frame rates, potentially making games visually indistinguishable from reality. The technology launches this fall and could reshape gaming experiences by making high-end visual quality accessible to more players.

Announcing NVIDIA DLSS 5, an AI-powered breakthrough in visual fidelity for games, coming this fall. DLSS 5 infuses pixels with photorealistic lighting and materials, bridging the gap between rendering and reality. Learn More → https://x.com/NVIDIAGeForce/status/2033617732147810782

Nvidia’s DLSS 5 delivers photorealistic neural rendering, shipping fall 2026
This represents the biggest graphics leap since real-time ray tracing, using AI to generate movie-quality lighting and materials in real-time gaming. The technology is already shifting industry sentiment, with former AI skeptics in the 3D community now defending neural rendering against critics. Early demos show the system running at 4K resolution and have impressed technical experts like Digital Foundry.

DLSS 5 is completely mind blowing. The neural rendering model with photoreal lighting and materials is a generation step up in visual fidelity. Gaming with DLSS 5 feels like future tech, but its possible now. It is truly incredible. 🤯 https://x.com/GeForce_JacobF/status/2033615891045454112

DLSS 5 might be the moment where the anti AI pendulum starts swinging back. Many in the 3D community who were against generative AI are now pushing back on the “”everything is AI slop”” crowd. The pendulum swung too far and they can feel it. Nice to see the rebalancing. https://x.com/bilawalsidhu/status/2034281398052274666

Here’s everything we know about Nvidia’s “”greatest leap in graphics since real-time ray tracing”” You can see Digital Foundry’s jaw drop in this reaction after they just saw DLSS 5.0: – Will ship in Fall of 2026! – Demo ran 4k on 2 5090’s but is already running on single GPU in https://x.com/Grummz/status/2033641075806769382

Palantir deploys Maven Smart System across entire Defense Department
The Pentagon’s Chief AI Officer demonstrated Palantir’s new software-as-a-service platform that will standardize AI decision-making tools across all military branches, marking the largest single AI deployment in Defense Department history and potentially reshaping how the military processes intelligence and makes operational decisions.

Palantir on X: “”This is Maven Smart System—Palantir’s software as a service product that we are deploying across the entire department.” https://t.co/hIaQAiq4iJ / X https://x.com/PalantirTech/status/2032142543022960980

Probably the most current look at Palantir’s maven smart system software. Here’s the DoW’s Chief AI officer showing how it works: https://x.com/bilawalsidhu/status/2032432668105712093

Niantic built global 3D map using 30 billion crowdsourced images
The Pokémon Go creator quietly collected massive visual data from players worldwide to construct detailed three-dimensional maps, raising concerns about how gaming companies harvest user-generated content for commercial mapping projects without explicit consent.

People are undoubtedly a little alarmed at having unwittingly helped build a 3D map of the world for Niantic by contributing 30 billion crowdsourced images. I interviewed Niantic’s CTO Brian McClendon about exactly this in a TED interview last year — he’s also the guy who https://x.com/bilawalsidhu/status/2033350363982471182

Google DeepMind launches framework to measure progress toward artificial general intelligence
The company released a cognitive science-based taxonomy identifying 10 key mental abilities needed for AGI, from perception to social cognition, and launched a $200,000 Kaggle competition to build evaluations that compare AI systems against human performance baselines. This represents the first systematic attempt to create empirical tools for tracking how close AI systems are to achieving general intelligence, moving beyond vague claims to measurable cognitive benchmarks.

Measuring Progress Towards AGI: A Cognitive Framework https://blog.google/innovation-and-ai/models-and-research/google-deepmind/measuring-agi-cognitive-framework/

Anthropic launches deployment arm to help businesses implement AI systems
Companies are rushing to integrate AI into their operations, and Anthropic is responding by creating a dedicated service to help over 1 million businesses deploy Claude and other AI tools. This move signals intense enterprise demand for practical AI implementation support, distinguishing Anthropic from competitors focused primarily on model development rather than hands-on deployment assistance.

This news came out a little earlier than we planned; we’re excited to be building a deployment arm and will share more details soon. Companies have a ton of urgency to deploy AI in their organizations and we’re sprinting to meet that demand. More than 1 million businesses run on https://x.com/fidjissimo/status/2033537381907710092

Tech consultant uses ChatGPT and AlphaFold to design dog cancer vaccine
Paul Conyngham, with no biology background, sequenced his rescue dog’s tumor DNA and fed it to AI tools to create a personalized mRNA vaccine that shrank the cancer by half. The case demonstrates how AI is democratizing complex scientific work, allowing non-experts to tackle problems previously requiring years of specialized training. A structural biologist called it “exciting” that someone outside the field could achieve such results.

An AI consultant with no biology training used ChatGPT and AlphaFold to create a personalized mRNA cancer vaccine for his rescue dog. Tumor shrunk by half. UNSW structural biologist Dr. Kate Michie: “It’s exciting to me that someone who’s not a scientist has been able to do https://x.com/TheRundownAI/status/2032843584869708105

this is actually insane > be tech guy in australia > adopt cancer riddled rescue dog, months to live > not_going_to_give_you_up.mp4 > pay $3,000 to sequence her tumor DNA > feed it to ChatGPT and AlphaFold > zero background in biology > identify mutated proteins, match them to https://x.com/IterIntellectus/status/2032858964858228817

How AI empowered Paul Conyngham to create a custom mRNA vaccine to cure his dog’s cancer when she had only months to live. The first personalized cancer vaccine designed for a dog: https://x.com/gdb/status/2032867435704103006

Florida homeowner sells house in 5 days using ChatGPT instead of agent
Robert Levine used AI to price, market, and draft contracts for his Cooper City home, receiving five offers within 72 hours and saving an estimated 3% in commission fees. The experiment demonstrates how AI tools can handle complex real estate transactions that traditionally require professional agents, though Levine still hired a lawyer for legal review. This case suggests AI could significantly disrupt the residential real estate industry by enabling direct sales.

Florida man sold his house in just 5 days after letting ChatGPT handle the entire process instead of a real estate agent The AI handled pricing, marketing, showings, and even helped draft the contract https://x.com/i/birdwatch/t/2032864183918690675?source=6

i mean this story is insane. man used chatgpt to sell his house in 5 DAYS. got 5 offers in 72 hours. no real estate agents. saved so much money doing it too. he used AI to: > price the house (researched neighboring properties for sale) > wrote up the legal contracts (saving https://x.com/cryptopunk7213/status/2033194801852567620?s=46

Man uses ChatGPT to sell his Cooper City home – NBC 6 South Florida https://www.nbcmiami.com/news/local/innovation-on-6/man-uses-chatgpt-to-sell-his-cooper-city-home-it-exceeded-our-expectations/3778919/

Claude Code launches channels for real-time messaging integration
Anthropic introduced channels that let users message Claude directly through Telegram, Discord, and iMessage while maintaining persistent coding sessions, enabling developers to collaborate with AI remotely without losing context. This bridges the gap between conversational AI and development workflows by allowing Claude to receive external events and respond through familiar messaging platforms while keeping the same coding session active.

Push events into a running session with channels – Claude Code Docs https://code.claude.com/docs/en/channels

Read about the feature and how to build your own channels for Claude Code on our docs! https://x.com/neilhtennek/status/2034762489951658190

Today we’re launching channels for Claude Code as an experimental feature! A few days ago, I was fed up that I couldn’t text Claude on the go like I would any of my friends. But those days are gone! Claude is saved in my contacts and I can keep shipping on the go. https://x.com/neilhtennek/status/2034762196576805123

Anthropic launches Dispatch, letting users assign tasks to Claude remotely and return to completed work
This marks a shift from chatbots to true AI assistants that work independently across devices. Users can message Claude from their phone to start tasks on their computer, then return later to find finished work waiting. Early users report it delivers on the long-promised vision of AI assistants that actually assist rather than just chat, potentially replacing riskier third-party automation tools.

We’re shipping a new feature in Claude Cowork as a research preview that I’m excited about: Dispatch! One persistent conversation with Claude that runs on your computer. Message it from your phone. Come back to finished work. To try it out, download Claude Desktop, then pair https://x.com/felixrieseberg/status/2034005731457044577

We’re shipping a new feature in Claude Cowork as a research preview that I’m excited about: Dispatch! One persistent conversation with Claude that runs on your computer. Message it from your phone. Come back to finished work. To try it out, download Claude Desktop, then pair https://x.com/felixrieseberg/status/2034005731457044577?s=12

After using it a bit, Claude Cowork Dispatch covers 90% of what I was trying to use OpenClaw for, but feels far less likely to upload my entire drive to a malware site. https://x.com/emollick/status/2034067677157679379

OpenClaw, Anthropic version. Basically: assign work from anywhere, come back to finished results. This is what “”AI assistant”” was always supposed to mean. https://x.com/fdaudens/status/2034080669119152238

AI skills evolve beyond simple text into complex folder systems
Developers are discovering that effective AI skills require entire file structures with scripts, assets, and data rather than just text prompts, fundamentally changing how we engineer AI context and capabilities.

Worth Reading: 1) What are Skills? They’re not just text files. They’re folders that can include scripts, assets, data, etc. A skill is a folder…think of the entire file system as a form of context engineering. 2) The Description Field Is For the Model The description field https://x.com/claude_code/status/2034335585339375855#m

LangChain AI development framework hits one billion downloads milestone
The popular toolkit for building AI applications reached this download threshold while joining NVIDIA’s coalition to develop open-source AI models, signaling massive developer adoption of AI agent-building tools. This milestone reflects the surge in businesses and developers creating custom AI applications beyond basic chatbots, with LangChain becoming a dominant infrastructure choice for connecting AI models to real-world data and workflows.

Announced in Jensen’s keynote today: LangChain frameworks have crossed 1B downloads. We’re excited to join the NVIDIA Nemotron Coalition to help shape the open models that power these agents. ➡️ Read the announcement: https://t.co/CWlbAzhlXy ➡️ Check out the docs: https://x.com/LangChain/status/2033788913937195132

OpenAI launches GPT-5.4 mini and nano models for faster coding
OpenAI released GPT-5.4 mini and nano, smaller AI models optimized for coding and computer tasks that run over twice as fast as previous versions. The models enable “subagents” – AI assistants that can handle specific tasks while preserving the main AI’s focus on complex planning and architecture. This represents a shift toward specialized AI workflows rather than single large models doing everything, potentially making AI assistance more practical for knowledge workers who need quick responses for routine tasks.

A knowledge-work platform built around GPT-5.4 Pro level intelligence would be really useful. The gap between other models and what Pro can do on complex intellectual work remains stark. I would love to have access in a Codex-like platform with shared file spaces, subagents, etc https://x.com/emollick/status/2033959257196966360

GPT-5.4 mini matters for subagents because it changes what feels worth handing off. The parent thread should hold the architecture, plan, and progress narrative. Fast subagents can explore the repo, check hypotheses, and preserve the parent thread’s limited attention. https://x.com/nickbaumann_/status/2034134875234832540#m

We’re introducing GPT-5.4 mini and nano, our most capable small models yet. GPT-5.4 mini is more than 2x faster than GPT-5 mini. Optimized for coding, computer use, multimodal understanding, and subagents. For lighter-weight tasks, GPT-5.4 nano is our smallest and cheapest https://x.com/OpenAIDevs/status/2033953815834333608

Subagents are now supported in Codex. They’re very fun and make it possible to get large amounts of work done *quickly*: https://x.com/gdb/status/2033757784437895367

AI agents now handle 80% of customer service calls while doubling satisfaction scores
Companies like Delta and Hertz are using AI concierges that know customer history and preferences to provide luxury-level service at mass scale. This represents a fundamental shift from episodic problem-solving to continuous, proactive customer relationships—potentially making every business operate like a high-end concierge service. Early results show both dramatic cost reductions and significantly higher customer satisfaction, suggesting AI may finally solve the internet-era tradeoff between scale and personalized attention.

The internet ruined customer service. AI could save it. | Andreessen Horowitz https://a16z.com/the-internet-ruined-customer-service-ai-could-save-it/

Anthropic sues federal government over AI restrictions in 2026 case
AI company Anthropic filed a lawsuit against the U.S. Department of War and multiple federal agencies, seeking to block government restrictions on AI development through temporary restraining orders and preliminary injunctions. The case represents a significant escalation in tensions between AI companies and federal regulators, with the government filing opposition documents that include redacted Department of War materials. This marks one of the first major legal challenges by an AI company against comprehensive federal AI oversight measures.

Opposition/Response to Motion – #96 in Anthropic PBC v. U.S. Department of War (N.D. Cal., 3:26-cv-01996) – CourtListener.com https://www.courtlistener.com/docket/72379655/96/anthropic-pbc-v-us-department-of-war/

Anthropic surveyed 81,000 Claude users across 159 countries about AI hopes and fears
The largest qualitative AI study ever conducted reveals users want professional excellence and personal transformation from AI, but worry about job displacement and losing human capabilities. Nearly one in five respondents prioritized using AI to handle routine tasks for more meaningful work, while concerns about unemployment and cognitive dependence appeared alongside optimism in the same individuals. The study’s scale—enabled by AI interviewing—provides unprecedented insight into how people actually experience AI versus abstract projections.

We invited Claude users to share how they use AI, what they dream it could make possible, and what they fear it might do. Nearly 81,000 people responded in one week—the largest qualitative study of its kind. Read more: https://x.com/AnthropicAI/status/2034302152945144166#m

What 81,000 people want from AI \ Anthropic https://www.anthropic.com/features/81k-interviews

Amazon founder Jeff Bezos seeks $100 billion for AI chip manufacturing fund
Bezos is reportedly raising capital to build AI semiconductor production facilities, addressing the critical shortage of specialized chips needed for training large language models. This represents one of the largest private investments in AI infrastructure to date, potentially reducing dependence on existing chip manufacturers like NVIDIA and creating new competition in the AI hardware supply chain.

Exclusive | Jeff Bezos in Talks to Raise $100 Billion for AI Manufacturing Fund – WSJ https://www.wsj.com/tech/jeff-bezos-aims-to-raise-100-billion-to-buy-revamp-manufacturing-firms-with-ai-618a3cfe

New OCR model achieves record accuracy across 90+ languages
Chandra OCR 2 scored 85.9% on industry benchmarks, surpassing previous text recognition systems while adding capabilities like handwriting detection, mathematical equation parsing, and automatic image captioning that could streamline document digitization across industries.

NEW SOTA OCR MODEL DROPPED Congrats to @VikParuchuri and team for releasing Chandra OCR 2! – 85.9% on olmocr bench, making it first place 🏆 – 90+ language support – 4B model – Full layout information – Extracts + captions images and diagrams – Strong handwriting, math, form, https://x.com/nathanhabib1011/status/2034565076963991910

ByteDance delays global launch of Seedance 2.0 after Hollywood legal threats
The TikTok parent company postponed its AI video generator’s international rollout following cease-and-desist letters from major studios like Disney, who accused ByteDance of stealing intellectual property after viral clips showed AI-generated footage of celebrities like Tom Cruise. The delay highlights how copyright concerns are becoming a major barrier for AI companies trying to scale video generation tools globally.

ByteDance reportedly pauses global launch of its Seedance 2.0 video generator | TechCrunch https://techcrunch.com/2026/03/15/bytedance-reportedly-pauses-global-launch-of-its-seedance-2-0-video-generator/

ByteDance Suspends Launch of Video AI Model After Copyright Disputes With Hollywood — The Information https://www.theinformation.com/articles/bytedance-suspends-launch-video-ai-model-copyright-disputes-hollywood

ByteDance launches SeedProteo for custom protein design from scratch
ByteDance’s new AI model can generate entirely new proteins atom-by-atom, potentially accelerating drug discovery and biotechnology applications. Unlike existing protein prediction tools that analyze known structures, SeedProteo creates novel proteins that don’t exist in nature, with a public server now available for researchers to test.

Here is SeedProteo, our latest diffusion-based model for de novo all-atom protein design from ByteDance Seed! Our server is now live — feel free to give it a try! https://x.com/SeedFold/status/2033515503839514771

New tool visualizes AI’s potential impact across 143 million US jobs
A GitHub research tool maps 342 occupations by size and uses AI prompts to score each job’s exposure to artificial intelligence disruption, revealing that digitally-native roles like software development and data analysis face the highest potential for AI-driven transformation. The interactive visualization draws from Bureau of Labor Statistics data and allows users to explore different metrics, though creators emphasize the AI exposure scores are rough estimates that don’t predict job elimination. The tool’s novel approach of using large language models to systematically evaluate occupational vulnerability provides a new lens for understanding AI’s economic implications across the entire job market.

US Job Market Visualizer https://karpathy.ai/jobs/

DoorDash couriers can now earn extra money training AI systems
The food delivery company launched a program where drivers complete paid tasks like transcribing audio and labeling images to help train artificial intelligence models. This represents a shift toward gig workers becoming direct contributors to AI development rather than just being replaced by it, with DoorDash leveraging its existing workforce of over one million couriers to generate training data for machine learning systems.

DoorDash’s New Paid Tasks Turn Couriers Into AI and Robot Trainers – Bloomberg https://www.bloomberg.com/news/articles/2026-03-19/doordash-s-new-paid-tasks-turn-couriers-into-ai-and-robot-trainers

Humanities education gains relevance as AI systems learn from human culture
As large language models train on humanity’s written works, understanding literature, history, and philosophy becomes crucial for effectively using AI tools, while providing essential context for navigating this technological transformation.

Never been a better time to study the humanities: 1) LLMs are trained on the cultural history of all humans, knowing that helps you use them 2) The humanities gives us context in this odd moment in history 3) Books & stuff are good Wrote this 3 years ago: https://x.com/emollick/status/2033692979844485304

ChatGPT builds playable strategy game entirely within Excel spreadsheet
A user tested three AI assistants on creating an Excel-based strategy game, with ChatGPT delivering the only fully functional result including game logic and an AI opponent using formulas. This demonstrates how different AI systems vary dramatically in their ability to translate creative requests into working solutions, even when using the same constrained platform like Excel.

Hey Excel agents from Claude, OpenAI & MS Copilot: “”make me a working strategy game in excel, it should have some form of graphics”” Claude made a board and acted as game master, Copilot created a board but no game, ChatGPT built a working game with formulas with a “”smart”” enemy. https://x.com/emollick/status/2033372471395512566

Google launches Stitch, an AI design tool that creates apps from speech
Google’s redesigned Stitch tool lets users create high-fidelity app interfaces using natural language and voice commands, then export them as functional React code. This represents a significant leap beyond typical design tools by bridging the entire pipeline from spoken concept to working software, positioning Google to compete directly with traditional design workflows. The tool features a 3D workspace, real-time voice collaboration with AI agents, and automatic prototype generation that can produce actual deployable applications rather than just mockups.

Design UI using AI with Stitch from Google Labs https://blog.google/innovation-and-ai/models-and-research/google-labs/stitch-ai-ui-design/

Exclusive: Early look at upcoming design tool from Google https://www.testingcatalog.com/exclusive-early-look-at-upcoming-vibe-design-tool-from-google/

Google expands Personal Intelligence feature across Search and Gemini apps
Google’s Personal Intelligence now connects data across Gmail, Photos, and other Google apps to provide personalized AI responses like custom shopping recommendations and travel itineraries. The feature launches for free U.S. users with user-controlled privacy settings, marking a shift from generic AI responses to deeply personalized assistance. Users can toggle app connections on or off, and Google says it doesn’t train directly on personal content like email inboxes.

Personal Intelligence in AI Mode and Gemini expands in the U.S. https://blog.google/products-and-platforms/products/search/personal-intelligence-expansion/

Google AI Studio launches full-stack coding agent for production apps
Google’s new Antigravity coding agent transforms simple prompts into complete web applications with multiplayer features, databases, and real-world API integrations, moving beyond basic prototypes to production-ready software. The platform now includes Firebase backend support, secure credential storage, and frameworks like Next.js, enabling developers to build complex apps like multiplayer games and collaborative tools without leaving the coding environment. Internal testing shows hundreds of thousands of apps already built using this system.

Introducing the all new vibe coding experience in @GoogleAIStudio, feating: – One click database support – Sign in with Google support – A new coding agent powered by Antigravity – Multiplayer + backend app support and so much more coming soon! https://x.com/OfficialLoganK/status/2034656376450908203

Vibe Code to production with Google AI Studio https://blog.google/innovation-and-ai/technology/developers-tools/full-stack-vibe-coding-google-ai-studio/

Andrej Karpathy’s AutoResearch lets AI agents autonomously run machine learning experiments overnight
Karpathy released AutoResearch, a system where AI agents independently modify training code, run 5-minute experiments, and iterate based on results—potentially completing ~100 experiments while researchers sleep. Early adopters scaled it to GPU clusters, achieving 910 experiments in 8 hours versus 96 sequential runs, demonstrating how autonomous AI research could accelerate discovery beyond human-paced iteration cycles.

A Guide to Andrej Karpathy’s AutoResearch: Automating ML with AI Agents | DataCamp https://www.datacamp.com/tutorial/guide-to-autoresearch

Karpathy’s Autoresearch is bottlenecked by a single GPU. We removed the bottleneck. We gave the agent access to our K8s cluster with H100s and H200s and let it provision its own GPUs. Over 8 hours: • ~910 experiments instead of ~96 sequentially • Discovered that scaling model https://x.com/skypilot_org/status/2034681533051855173

karpathy/autoresearch: AI agents running research on single-GPU nanochat training automatically https://github.com/karpathy/autoresearch

LangChain releases plug-and-play AI agent system for developers
LangChain open-sourced Deep Agents, a pre-built system that gives developers working AI agents without manual setup of prompts and tools. This matters because it removes technical barriers that previously required extensive coding to create functional AI assistants. The MIT license ensures broad commercial adoption, potentially accelerating AI agent deployment across businesses.

LangChain just open-sourced Deep Agents—an agent harness that’s opinionated and ready-to-run out of the box. Instead of wiring up prompts, tools, and context management yourself, you get a working agent immediately and customize what you need. It’s an MIT-licensed system that’s https://x.com/itsafiz/status/2033591253955449289

LangChain launches enterprise agent management platform LangSmith Fleet
The platform lets companies build AI agents using plain English, then deploy them across teams with built-in security controls, human approval workflows, and action tracking—addressing the key barriers that have kept AI agents from widespread business adoption.

Introducing LangSmith Fleet: an enterprise workspace for creating, using, and managing your fleet of agents. Fleet agents have their own memory, access to a collection of tools and skills, and can be exposed through the communication channels your team uses every day. Fleet https://x.com/LangChain/status/2034679590250258855

Introducing LangSmith Fleet. Agents for every team. → Build agents with natural language → Share and control who can edit, run, or clone each agent → Manage authentication with agent identity → Approve actions with human-in-the-loop → Track and audit actions with tracing in https://x.com/LangChain/status/2034694530478612777

Enterprise AI teams rediscover MCP after CLI hype cycle fades
While individual developers save tokens using command-line interfaces, enterprises are returning to Model Context Protocol for structured agent deployment at scale. The shift highlights a key distinction: CLI tools work well for solo coding but lack the authentication, monitoring, and standardization that organizations need when multiple teams use AI agents across sensitive systems.

@chrlschn – MCP is Dead; Long Live MCP! https://chrlschn.dev/blog/2026/03/mcp-is-dead-long-live-mcp/

good concepts here in fastmcp on distributing skills via MCP resources, i think this might be the right approach solves the problem of skills going out of date if you load them fresh each time + can tie them to tools easier more to explore here https://x.com/RhysSullivan/status/2034125767987368242#m

MCP was a mistake. Long live CLIs. https://x.com/skirano/status/2034269154404868314#m

Meta plans 20% workforce cuts as AI investments reach $600 billion
The social media giant is preparing its largest layoffs since 2023 to offset massive AI infrastructure spending, with CEO Zuckerberg citing efficiency gains where single employees can now accomplish work that previously required entire teams. The cuts reflect a broader tech industry trend of using AI capabilities to justify significant workforce reductions, following similar moves by Amazon and Block.

Exclusive: Meta planning sweeping layoffs as AI costs mount | Reuters https://www.reuters.com/business/world-at-work/meta-planning-sweeping-layoffs-ai-costs-mount-2026-03-14/

Microsoft launches Copilot Health to analyze personal medical data
Microsoft’s new Copilot Health platform integrates health records, wearable data, and lab results to provide personalized medical insights, aiming to help patients prepare better questions for doctors rather than replace medical care. The service connects to over 50,000 US hospitals and 50+ wearable devices, representing a significant step toward what Microsoft calls “medical superintelligence” – AI that combines general physician knowledge with specialist expertise. Early access requires joining a waitlist, with the platform launching first in the US for adults 18 and older.

I’m not saying Copilot diagnosed me. I’m saying it helped me ask for the right test. A test no doctor had ordered for me in twenty years.”” This is why I’m so passionate about AI & healthcare. https://x.com/mustafasuleyman/status/2033655842919395723

Introducing Copilot Health | Microsoft AI https://microsoft.ai/news/introducing-copilot-health/

We’re approaching the dawn of medical superintelligence – the moment when affordable, world-class medical knowledge and support is at your fingertips whenever you need it. I think people are still underestimating how profound this transformation is going to be. Today we’re https://x.com/mustafasuleyman/status/2032092644483141928

Microsoft creates virtual cancer populations to accelerate tumor research
Microsoft developed AI that generates synthetic cancer tissue samples from standard $5 biopsy slides, potentially replacing expensive specialized tests that cost hundreds of dollars. The system creates virtual patient populations to help researchers study tumor environments without needing rare or costly tissue samples. This could dramatically speed up cancer research by making advanced tissue analysis accessible to any lab with basic microscopy equipment.

Every time you get a cancer biopsy, the lab makes a tissue slide that costs about $5. It shows the shape of your cells under a microscope, and every cancer patient already has one on file. There’s a much fancier version of that test called multiplex immunofluorescence (basically https://x.com/anishmoonka/status/2033344818475360562

GigaTIME: Scaling tumor microenvironment modeling using virtual population generated by multimodal AI – Microsoft Research https://www.microsoft.com/en-us/research/blog/gigatime-scaling-tumor-microenvironment-modeling-using-virtual-population-generated-by-multimodal-ai/

Midjourney launches V8 model with 5x speed boost and better prompt following
The AI image generator’s latest version delivers significantly faster generation times while improving core capabilities like text rendering and style personalization. This represents a major performance leap for one of the leading AI art platforms, potentially making high-quality image generation more accessible through reduced wait times and better user control over creative outputs.

Today we’re starting to test an early version of our V8 model with our community. It’s much better at following prompts, 5x faster, has native 2K modes, improved text rendering and the best personalization, sref, and moodboard performance ever. Have fun! https://x.com/midjourney/status/2034015403542974793?s=20

MiniMax releases M2.7, first AI model to help build itself
MiniMax’s M2.7 model achieved breakthrough “self-evolution” by autonomously running over 100 optimization loops during its own development, scoring 56.22% on advanced coding benchmarks and reducing production incident recovery times to under three minutes. This represents the first commercially available AI that meaningfully participated in improving its own training process, moving beyond human-only model development toward AI systems that can enhance themselves.

BREAKING 🚨: MiniMax released MiniMax M2.7, a new self-evolving model, achieving a score of 56.22% on SWE-Bench Pro. M2.7 was used for building complex agent harnesses during its own development. Users can now access MiniMax M2.7 via APIs and MiniMax Agent. https://x.com/testingcatalog/status/2034250919345377604#m

During the iteration process, we also realized that the model’s ability to recursively evolve its harness is equally critical. Our internal harness autonomously collects feedback, builds evaluation sets for internal tasks, and based on this continuously iterates on its own https://x.com/MiniMax_AI/status/2034315323109953605#m

Introducing MiniMax-M2.7, our first model which deeply participated in its own evolution, with an 88% win-rate vs M2.5 – Production-Ready SWE: With SOTA performance in SWE-Pro (56.22%) and Terminal Bench 2 (57.0%), M2.7 reduced intervention-to-recovery time for online incidents https://x.com/MiniMax_AI/status/2034315320337522881#m

MiniMax Global Announces Full Year 2025 Financial Results – MiniMax News | MiniMax https://www.minimax.io/news/minimax-global-announces-full-year-2025-financial-results

Minimax M2.7 released! And its a big one Highlights: Self-evolving – first model that helped build itself, running 100+ autonomous optimization loops during its own RL training (30% internal improvement). Strong coder – 56.2% on SWE-Pro (near Opus 4.6), 55.6% on VIBE-Pro, https://x.com/kimmonismus/status/2034269026353082422#m

MiniMax M2.7: Early Echoes of Self-Evolution – MiniMax News | MiniMax https://www.minimax.io/news/minimax-m27-en

Mistral launches Forge for enterprises to train custom AI models
Mistral AI unveiled Forge, a system that lets companies train frontier-grade AI models on their proprietary data rather than generic public datasets. This addresses a key enterprise limitation where standard AI models lack understanding of internal processes, compliance policies, and institutional knowledge. Major organizations including ASML, Ericsson, and the European Space Agency are already using Forge to create models that understand their specific terminology, workflows, and operational constraints. Mistral releases first open-source Lean 4 code agent for formal verification Mistral AI introduced Leanstral, the first open-source AI agent designed for Lean 4 formal proof systems, enabling automated mathematical proofs and software verification. This breakthrough addresses the bottleneck of human review in high-stakes coding by allowing AI to both generate code and formally prove its correctness. With only 6 billion active parameters, Leanstral outperforms much larger models while costing 92 times less than Claude Opus, potentially transforming how critical software and mathematical research is developed and verified.

Introducing Forge | Mistral AI https://mistral.ai/news/forge

Leanstral: Open-Source foundation for trustworthy vibe-coding | Mistral AI https://mistral.ai/news/leanstral

Introducing Mistral Small 4 | Mistral AI https://mistral.ai/news/mistral-small-4

Kimi introduces attention residuals to replace fixed connections in AI models
Chinese AI company Kimi released a new technique called “attention residuals” that replaces the standard fixed connections between layers in transformer models with learnable attention mechanisms. This could improve how AI models process information across their many layers, moving beyond the uniform accumulation method used since transformers were invented. The research is generating significant discussion in the AI community, with experts calling it a fundamental rethinking of how deep neural networks aggregate information.

@_avichawla Impressive work from Kimi https://x.com/elonmusk/status/2033528245464047805

🔥 @Kimi_Moonshot’s new Attention Residual paper is sparking discussions. Zhihu contributor OpenLLMAI shares a deep dive: “”From Kimi’s Attention Residual to ‘Vertical Attention’ — an idea I’ve been thinking about for half a year.”” Some interesting thoughts on attention mechanisms https://x.com/ZhihuFrontier/status/2033751367198949865

Avi Chawla on X: “Big release from Kimi! They just released a new way to handle residual connections in Transformers. In a standard Transformer, every sub-layer (attention or MLP) computes an output and adds it back to the input via a residual connection. If you consider this across 40+ layers, https://t.co/5i5AN9tzIm / X https://x.com/_avichawla/status/2033472650836914495

Introducing 𝑨𝒕𝒕𝒆𝒏𝒕𝒊𝒐𝒏 𝑹𝒆𝒔𝒊𝒅𝒖𝒂𝒍𝒔: Rethinking depth-wise aggregation. Residual connections have long relied on fixed, uniform accumulation. Inspired by the duality of time and depth, we introduce Attention Residuals, replacing standard depth-wise recurrence with https://x.com/Kimi_Moonshot/status/2033378587878072424

visual summary of attention residuals by kimi, beautiful paper https://x.com/eliebakouch/status/2033488233854620007

Hermes Agent gains traction as simpler alternative to OpenClaw for AI automation
Multiple developers report that Hermes Agent, an AI automation platform from Nous Research, offers easier setup and more reliable performance than competing tool OpenClaw. Users highlight its out-of-the-box functionality, superior memory systems, and ability to handle complex tasks like geopolitical prediction modeling and media server management without requiring frequent manual intervention. The platform’s recent v0.3.0 update added real-time streaming and a plugin architecture for sharing custom tools.

Built PrediHermes ✨ a Hermes Agent skill + companion WorldOSINT/MiroFish forks for geopolitical prediction. @NousResearch It pulls 54+ OSINT modules, uses Polymarket to find contracts with clear resolution criteria, then runs MiroFish multi-agent sims to model individual https://x.com/WeXBT/status/2033391568426598608

Hermes Agent v0.3.0 ☤ 248 PRs. 15 contributors. 5 days. • Real-time streaming across CLI and all platforms • First-class plugin architecture, package and share tools+commands+skills • /browser connect to live Chrome via CDP • @vercel AI Gateway model provider • https://x.com/NousResearch/status/2033877040399831478

Browser Use is now an official provider for the browser tool in Hermes-Agent – Update to try it out 😉 Use `hermes tools` to set the browser backend. (Note: this requires an API key with them) https://x.com/Teknium/status/2033811117521408078

Did a small local anime server tool powered by Hermes Agent (@NousResearch). You can: – fully sync your anime list – download torrents from different sources – add tracking & scheduled downloads – auto-manage disk usage – serve to any device within your local wifi and more! https://x.com/rodmarkun/status/2033307437088850102

I have consecutively spent millions of tokens today with Hermes without breaking anything, where Openclaw would’ve needed several nudges. Both have their merits, but for prod, Hermes nailed it. Amazing job @Teknium https://x.com/populartourist/status/2034653545287348266

i’ve been using @NousResearch Hermes Agent for about a week and my initial thoughts are it just works out of the box I find it’s memory and learning to be far superior to OpenClaw without augmenting it with QMD or any additional memory systems it’s early and my OC’s handle A https://x.com/austin_hurwitz/status/2033552632241857002

Told my agent to create a fresh OpenClaw agent in digital ocean according to the instructions. Ran into like 6 issues with the flow from the official docs. Did the same for the @NousResearch Hermes agent – Claude one-shotted it 🙂 https://x.com/0xMasonH/status/2033608276286243323

trying Hermes agent. setup was significantly easier than openclaw. I can see its already basically a leaner and more opinionated openclaw https://x.com/fuckyourputs/status/2033503910376431728

Healthcare robots learn from massive new dataset of medical tasks
Researchers released the first comprehensive dataset of healthcare robot movements, training AI models to perform medical procedures like drawing blood and patient positioning. This breakthrough could accelerate deployment of robots in hospitals by giving them foundational skills learned from thousands of recorded medical tasks, potentially addressing healthcare worker shortages. The dataset represents a shift from general-purpose robotics to specialized medical AI that understands the unique requirements of patient care.

The First Healthcare Robotics Dataset and Foundational Physical AI Models for Healthcare Robotics https://huggingface.co/blog/nvidia/physical-ai-for-healthcare-robotics

Nvidia delivers first DGX Station GB300 to Karpathy’s lab
Former Tesla AI chief Andrej Karpathy received Nvidia’s latest high-end AI workstation, signaling continued investment in independent AI research. The DGX Station GB300 represents Nvidia’s newest desktop supercomputer designed for AI development, suggesting major computational advances may emerge from Karpathy’s new laboratory work.

🙌 Andrej Karpathy’s lab has received the first DGX Station GB300 — a Dell Pro Max with GB300. 💚 We can’t wait to see what you’ll create @karpathy! 🔗 https://t.co/8ct5QZ3frS @DellTech https://x.com/NVIDIAAIDev/status/2034291235041554871

Nvidia CEO calls future computers “token manufacturing systems” at GTC 2026
Jensen Huang outlined Nvidia’s vision where computers primarily generate AI tokens rather than run traditional software, while unveiling GR00T N2, a new robotics foundation model that uses integrated world models instead of vision-language approaches. This represents a fundamental shift from general computing to specialized AI processing infrastructure, with Nvidia positioning itself as both vertically integrated across the AI stack and horizontally open to partners.

What is computer in the future? According to Jensen: “In the future, the computer is really a manufacturing system for tokens. And the number of computers in the world built for token manufacturing is still very small. It’s small because most of the systems we have shipped so https://x.com/TheTuringPost/status/2033983885131059636

Jensen: “Nvidia is the first vertically integrated but horizontally open company.” This strategy positions Nvidia as the backbone of robotics without stifling innovation. Vertical integration ensures cutting-edge performance on each layer of the AI stack. Horizontal openness https://x.com/TheHumanoidHub/status/2033622691408974133

GR00T is moving away from VLM-based backbones in favor of integrated world models. Jensen Huang teased GR00T N2 during his keynote; NVIDIA’s next-gen foundation model built on DreamZero research. Utilizing a new world-action model architecture, it succeeds at novel tasks in https://x.com/TheHumanoidHub/status/2034279221372321940

NVIDIA GTC 2026 Keynote: Everything That Happened in 12 Minutes – YouTube https://www.youtube.com/watch?v=X2i_8O75_Os

Runway and Nvidia achieve instant HD video generation in under 100ms
The companies developed a real-time AI video model that generates high-definition footage with near-zero delay, eliminating the lengthy processing times that have limited AI video tools. This breakthrough could transform live streaming, gaming, and interactive media by making AI video generation as responsive as traditional cameras.

A breakthrough in real-time video generation. As a research preview developed with @NVIDIA and shared at @NVIDIAGTC this week, we trained a new real-time video model running on Vera Rubin. HD videos generate instantly, with time-to-first-frame under 100ms. Unlocking an entirely https://x.com/runwayml/status/2034284298769985914#m

Stanford and NVIDIA create robot that predicts physical world from photos
PointWorld-1B uses a single image plus depth data to simulate how entire 3D scenes will move when a robot takes actions, essentially giving machines the ability to mentally rehearse tasks before performing them. This breakthrough could dramatically improve robot planning and safety by letting them predict consequences in complex real-world environments. The model represents a significant leap from current AI that typically processes static images to dynamic world understanding.

What if a robot could simulate the physical world from a single image. [📍Bookmark Paper & GitHub for later] PointWorld-1B from Stanford and NVIDIA is a large 3D world model that predicts how an entire scene will move, given RGB-D input and robot actions. The key idea is https://x.com/IlirAliu_/status/2032895393407660380

OpenAI releases GPT-5.4 mini matching larger model performance at lower cost
OpenAI launched GPT-5.4 mini, a smaller model that nearly matches its full-sized counterpart on coding benchmarks while using 30% less computing quota in development tools. The mini version costs $0.75 per million input tokens versus the previous mini’s pricing, representing a 2.25x increase but still offering significant savings for simpler tasks. This release suggests OpenAI is focusing on efficiency improvements rather than dramatic capability jumps, with some observers noting the incremental nature warranted a point release rather than a full version upgrade to GPT-6.

5.3 to 5.4 is what i would have expected to warrant a jump to GPT-6 https://x.com/yacineMTB/status/2033291560217923803

GPT-5.4 mini approaches the performance of the larger GPT-5.4 model on several evaluations, including SWE-Bench Pro and OSWorld-Verified. https://x.com/OpenAIDevs/status/2033953828387885470

GPT-5.4 mini is available today in the API, Codex, and ChatGPT. In the API, it has a 400k context window. In Codex, it uses only 30% of the GPT-5.4 quota, letting you handle simpler coding tasks for about one-third of the cost. GPT-5.4 nano is only available in the API. https://x.com/OpenAIDevs/status/2033953840312291603

GPT-5.4-mini 2.25 times more expensive than GPT-5-mini $0.75 Input $4.5 Output 400k https://x.com/scaling01/status/2033955279079907511

Introducing GPT-5.4 mini and nano | OpenAI https://openai.com/index/introducing-gpt-5-4-mini-and-nano/

OpenAI acquires Python tooling startup Astral for developer infrastructure
OpenAI is buying Astral, the company behind popular Python development tools like Ruff and uv, signaling a strategic push beyond AI models into the software development toolchain itself. This acquisition suggests OpenAI wants to control more of the programming workflow, potentially integrating AI assistance directly into the coding tools that millions of developers use daily. The move positions OpenAI to compete with Microsoft’s GitHub Copilot by owning the underlying development infrastructure rather than just providing AI features on top of existing tools.

OpenAI to acquire Astral | OpenAI https://openai.com/index/openai-to-acquire-astral/

OpenAI partners with Amazon to sell AI tools to US government agencies
OpenAI signed a deal with Amazon Web Services to distribute its AI models to federal agencies through AWS’s government cloud infrastructure, including classified networks. This move directly challenges Anthropic’s dominance in the government AI market, where Amazon has invested $4 billion and deeply integrated Claude models into AWS services. The partnership expands OpenAI’s federal reach beyond its recent Pentagon contract and could boost its credibility with enterprise customers who view government approval as a trust signal.

OpenAI expands government footprint with AWS deal, report says | TechCrunch https://techcrunch.com/2026/03/17/openai-expands-government-footprint-with-aws-deal/

Encyclopedia Britannica and Merriam-Webster sue OpenAI for copyright infringement
The publishers claim OpenAI illegally scraped nearly 100,000 articles to train its AI models and that ChatGPT directly competes with their content by reproducing it without permission. This lawsuit is distinctive because it challenges both the training process and ChatGPT’s real-time web search feature that pulls content from publishers’ sites. The case joins a growing wave of legal challenges from major publishers, with legal precedent still unclear on whether using copyrighted content for AI training constitutes fair use.

The dictionary sues OpenAI | TechCrunch https://techcrunch.com/2026/03/16/merriam-webster-openai-encyclopedia-brittanica-lawsuit/

OpenAI pivots ChatGPT to enterprise productivity ahead of 2026 IPO
The AI giant is “orienting aggressively” toward business use cases to convert its 900 million weekly users into higher-paying enterprise customers, with a public debut planned for Q4 2026. This marks a strategic shift from consumer chat to workplace productivity tools, as OpenAI projects $280 billion revenue by 2030 split equally between consumer and enterprise segments.

OpenAI preps for IPO in 2026, says ChatGPT must be ‘productivity tool’ https://www.cnbc.com/2026/03/17/openai-preps-for-ipo-in-2026-says-chatgpt-must-be-productivity-tool.html

Walmart’s ChatGPT shopping partnership shows disappointing conversion rates
Walmart’s experiment letting customers shop through ChatGPT, with OpenAI earning commissions on purchases, has yielded poor results with low conversion rates from browsing to buying. This suggests that while AI chatbots excel at conversation, they may struggle to replicate the visual and interactive elements that drive actual e-commerce purchases, highlighting a key limitation in AI’s retail applications.

$WMT is disappointed in results from OpenAI partnership, whereby Walmart users are allowed to shop via ChatGPT and OpenAI would receive a commission on these purchases “Conversion rates—the percentage of users following through with a purchase of an item shown to them by https://x.com/negligible_cap/status/2034369496543305971?s=46

GPT-4 tutoring system boosts high school test scores significantly
A randomized study found that personalized AI tutoring raised student test scores by an amount equivalent to six to nine months of additional schooling. This represents the first rigorous experimental evidence that AI can meaningfully improve educational outcomes, moving beyond theoretical potential to measurable academic gains.

AI really can help education: Randomized controlled experiment on high school students found a GPT-4o powered tutor that personalized problems for students raised final test scores by .15 SD, “”equivalent to as much as six to nine months of additional schooling by some estimates”” https://x.com/emollick/status/2033773791688433708

OpenClaw beta adds live browser control through Chrome’s remote debugging
This AI tool can now directly manipulate web browsers in real-time, marking a shift from AI that just reads websites to AI that can actively control them. The integration with Chrome’s debugging features suggests AI agents are moving toward more autonomous web interactions, potentially changing how we think about browser automation and digital task completion.

New @openclaw beta is up: it comes with the new live browser control that Google added in latest Chrome! enable via chrome://inspect#remote-debugging Your clanker will know when to use what, or you can ast it. new “”user”” profile session is there! https://x.com/steipete/status/2032686376932491363

Computer launches browser agent that controls any website or app
Anthropic’s AI assistant can now navigate the web independently through users’ browsers, eliminating the need for custom integrations to access online services. This represents a significant leap from text-based AI to autonomous web interaction, potentially transforming how people complete digital tasks by having AI directly manipulate websites and applications on their behalf.

Computer can now take full control of Comet to complete tasks. When you’re in Comet, Computer spins up a browser agent that can access any site or logged‑in app with your permission, without the need for connectors or MCPs. Available to all Computer users on Comet. https://x.com/perplexity_ai/status/2033598416962592813

Perplexity launches health AI platform with data dashboards and specialized agents
Perplexity Health combines personal health data visualization with purpose-built AI agents for nutrition and sleep, differentiating it from competitors like Microsoft’s Copilot Health and OpenAI’s ChatGPT Health that focus mainly on medical Q&A. The platform follows Perplexity’s proven strategy of connecting real user data with AI reasoning, similar to its successful Finance product that now accesses over 40 live financial tools. This marks another expansion for the $21 billion company as it transforms from a search engine into a comprehensive personal intelligence platform.

Perplexity launches Perplexity Health agent in US https://www.testingcatalog.com/perplexity-launches-perplexity-health-agent-in-us/

Introducing Perplexity Health https://www.perplexity.ai/hub/blog/introducing-perplexity-health

AI-generated Japanese metal band fools 80,000 Spotify listeners
A completely fake band called Neon Oni, created using Suno AI music generation, attracted thousands of genuine fans who included the non-existent group in their top Spotify playlists before being exposed by online investigators. The incident demonstrates how AI can now create convincing musical acts complete with backstories and merchandise, blurring the line between authentic and artificial entertainment in ways that fool even dedicated music fans.

Someone used Suno AI to generate a Japanese metal band called Neon Oni. Fake member bios, AI-generated music videos, “”Based in Tokyo”” on Spotify. 80,000+ monthly listeners. Fans had it in their Spotify Wrapped top 5. Merch was selling. Then, community sleuths exposed it. Traced https://x.com/TheRundownAI/status/2033568236227244451?s=20

Schibsted open sources AI tool that converts news articles into videos
Norwegian media giant Schibsted released Videofy as open source software, allowing any newsroom to automatically transform written articles into broadcast-ready videos within minutes. The tool handles the entire production pipeline—generating scripts, selecting visuals, creating voiceovers, and assembling final videos—addressing a major resource constraint for smaller news organizations that lack video production capabilities. After producing thousands of videos internally, Schibsted aims to democratize video news creation and accelerate industry-wide innovation in AI-powered journalism tools.

Schibsted open sources AI tool that turns news articles into videos | Schibsted https://schibsted.com/news/schibsted-open-sources-ai-tool-that-turns-news-articles-into-videos/

Leave a Reply

Trending

Discover more from Ethan B. Holland

Subscribe now to keep reading and get access to the full archive.

Continue reading