About This Week’s Covers

This week’s theme celebrates one of my favorite Spring traditions: when I run around the mountains in the McDowell Sonoran Preserve outside Scottsdale, Arizona. It’s a challenging desert run that’s about nine miles with around 2,000 vertical feet of elevation gain.

When you see this from your hotel… the mountain calls you to run over and around it!

Four years ago, I attempted it and bailed. I got to the halfway point and turned around because I didn’t feel comfortable going around the backside of the McDowell mountains with no cell service, in the desert, in the heat, alone.

The second year, I ran the other direction halfway and chickened out again. I didn’t like the mystery of the backside with no cell phone, and even though mountain lions are shy, I didn’t like the idea of them watching me.

Last year was the first time I finally made it around the entire loop. I went counterclockwise, up to Bell Pass, then around the back of McDowell Peak to Wingate Pass, and then back down. It was a great feeling to finish the loop and come back unscathed, especially mentally getting past my fear while also being prepared with enough fluids.

This year, I decided to go clockwise, so I went up to Inspiration Viewpoint, across Wingate Pass, and then around the back of McDowell Peak and up through Bell Pass back to the trailhead.

I was able to get a photo of the trail from the airplane leaving Phoenix

It’s been a fun study in doing things alone and building comfort in new places. When we do something potentially dangerous alone for the first time, I think it’s always good to err on the side of caution. As we grow comfortable, improve, and become familiar, the mysteries gets less scary and simply turn into fun adventures.

And as luck would have it, this time I saw three people. Usually, I don’t see anybody on the backside, but this time I saw three different hikers. I also saw all sorts of animal poop, which was pretty cool, including fresh (!) mountain lion, bobcat, coyote, and great horned owl evidence.

I had Claude Cowork compile the best guesses from GPT, Gemini, and Claude for each photo, and then conclude the animal that each one came from… then GPT Images made me these nifty charts.

I head to Scottsdale each May for the Local Media Consortium‘s annual summit and board meeting. The LMC is one of my favorite organizations, and I’m honored to serve on the board of directors. It’s a four-day event, and during one of the afternoons, after the sessions, I skip the cocktail party and knock out a nine-mile jog in the mountains before dinner.

Going at the end of the day means it’s hot out in the desert. One year, it was in the high 90s and really tough. The entire trail is exposed, with no shade. This year, the temperatures were in the high 70s, a minor miracle.

For this week’s cover, I gave one of the actual images I took on the trail to ChatGPT, as well as Google Gemini. I ended up liking GPT’s image the best because the placement of the sign was nicely composed against the image.

Some of the runner ups for the cover… actual photos with signs added to them.

For this week’s category covers, I gave a photo of one of the trailhead signs to my Claude skill, along with a picture of a trail itself.

Claude takes my open-ended description of my theme and generates a JSON file, which then gets turned into prompts by Claude and run through the Gemini API. All I have to do is dictate a very open-ended description of my theme, and then my Python scripts take it from there. Ten minutes later I have 60 category covers.

I don’t this it as a replacement for making a creative effort. I do it because I want to learn how close to parity I can get with automation. I like that the category covers reference actual photos I took on the trail. They’re not really fully diffused images as much as in-painted with signs and objects on top of the image I took.

Here are a few of my favorite covers:

AGI has question marks about where the summit is because it’s limitless.
Audio has a little bird, and the names of the trails are Whisper, Waveform, and Podcast. Autonomous has a car driving, and the trail names refer to the topic.
Benchmarks has some actual deep-track benchmarks as the trail names, but I love that the elements it included are an old-fashioned distance-measuring wheel and a stopwatch.
I like that the Google image simply has a small ribbon that lets you find your place with the Google iconography. It’s a neat touch.
Images has very creative trail names, but it also has a picture of a portrait of the actual image itself, which is pretty recursive, accurate, and well-made. Props to Gemini for the accuracy on that one.
Inflection is neat because it inserted a fork in the road in my trail: one with an old course and one with a new direction that’s infinite.
The Internet has some creative trail names, as well as an old dusty modem and a rusty Wi-Fi logo.
Perplexity has questions for all of the trail markers. That’s pretty hilarious. Plus, the coyote looks inquisitively lost.
Publishing has great trail names, with a printing press icon as well as a newspaper lying at the bottom of the sign.
Qwen is great, with a small stone cairn and a sleeping panda bear hiding in the shade under a rock.
RAG has great trail names, as well as a little old library index-card shelf.
Robots has great trail names, plus a nice dirty version of the Boston Dynamics Spot robot.
Sakana, even though I never explained what it is, somehow the image tool knows that Sakana means fish and it’s a Japanese model, so they put a little koi pond at the top of the trail. Very creative.
And then World Models has a little terrarium in a jar of the desert, with miniature cacti inside it. Super creative.

Humanities Reading for The Week

For this week’s humanities reading, I gave Claude and GPT a description of my desert experiences, and I got some great recommendations. I do this not because I can’t find my own humanities readings (I was an English Major at Swarthmore College with a concentration in Black Studies), but because I want to teach myself how to build systems that can automatically generate ideas I vouch for, personally. And to be honest, the recommendations are solid.

The winner this week is a classic poem by Rainer Maria Rilke called “The Man Watching”.

The runners-up this week are worth shouting out:

I especially liked the fact that Rainer Maria Rilke popped up, because Rilke is a favorite of my good friend Alexis Deise. Alexis is an absolutely incredible quilter who crafts museum-grade-quality pieces. She incorporates a blend of advocacy, nature, history, study, and talent in her quilting. Her most recent quilt includes a great blue heron, which is one of our resident birds at the beach. The quilt is on currently display in Wisconsin. Alexis’ quilts are fantastic organic medicine.

Visit Alexis at https://www.instagram.com/alexisdeise

Another inspiration this week, that ties to the week’s wilderness theme, is my friend John Bayalis, who has also found peace by going out into nature. In particular, John goes backpacking in the Channel Islands for a week all by himself once a year and goes on a vision quest.

The Man Watching by Rainer Maria Rilke

I can tell by the way the trees beat, after
so many dull days, on my worried windowpanes
that a storm is coming,
and I hear the far-off fields say things
I can’t bear without a friend,
I can’t love without a sister.

The storm, the shifter of shapes, drives on
across the woods and across time,
and the world looks as if it had no age:
the landscape, like a line in the psalm book,
in seriousness and weight and eternity.

What we choose to fight is so tiny!
What fights with us is so great!
If only we would let ourselves be dominated
as things do by some immense storm,
we would become strong too, and not need names.

When we win it’s with small things,
and the triumph itself makes us small.
What is extraordinary and eternal
does not want to be bent by us.
I mean the Angel who appeared
to the wrestlers of the Old Testament:
when the wrestlers’ sinews
grew long like metal strings,
he felt them under his fingers
like chords of deep music.

Whoever was beaten by this Angel
(who often simply declined the fight)
went away proud and strengthened
and great from that harsh hand,
that kneaded him as if to change his shape.
Winning does not tempt that man.
This is how he grows: by being defeated, decisively,
by constantly greater beings.
German; trans. Robert Bly

https://www.ronnowpoetry.com/contents/rilke/ManWatching.html

This Week By The Numbers

Total Organized Headlines: 699

This Week’s Executive Summaries

This week I organized 699 links and 149 of them contributed to the executive summaries.

I’m organizing the top stories so that the biggest ones are first, and then the rest will be organized alphabetically by company or topic name.

I gave a speech to the Associated Press recently where I got to recap the last 16 months. Every week, for 16 months in a row, has been full of enough stories for entire year. There are probably 20 stories in any given week that would merit sitting down for an hour and unpacking.

The top story this week is that Anthropic saw its revenue grow 80 times on an annualized basis as of the first quarter of 2026. They were planning 10x growth, but ended up seeing 80x. This has led to an incredible spree of spending as Anthropic tries to cover the amount of computing power they need to handle this crazy growth.

Two of the big computing spends merit a note: First, Anthropic is going to spend $200 billion on Google Cloud computing; and second, Anthropic is going to rent the entire Colossus 1 data center from SpaceX. That’s 220,000 NVIDIA GPUs… from SpaceX.

This is a tell that Grok doesn’t need the power, which is a whole other story. SpaceX buying xAI, to me, is a continued obsfucation game where Elon is hiding his losses, which initially started from his massive purchase of Twitter. Twitter hid it by laundering it into xAI, and now that’s being swept under the rug as it is acquired by SpaceX and recouped by SpaceX’s IPO at an insane multiple of 56x to 110x revenues.

A recently as a few weeks ago, Elon was lashing out at Amanda Askell and the Anthropic team, with hyperbolic insults.

Then, Elon pivoted…

In other Anthropic news, Anthropic launched 10 pre-built Claude agents for financial services. This is a big deal because so many financial AI services are basically frontier model wrappers with no moat whatsoever. Quite a bit of “software” is just an agent markdown file.

So if Anthropic releases these pre-built agentic systems that you can add in as a plug-in to Claude… everything’s just a prompt away, and you might as well use the prompt that’s included with your $20-a-month Claude fee.

Not all Claude’s financial agents are headline makers, but they’ve got strong ones like earnings reviewers and market researchers. They’ve got a general ledger reconciler and a month-end closer. They also have a statement auditor that reviews statements for consistency and whether or not they’re ready to be audited.

This is similar to another set of agentic skills that Anthropic released for nonprofits just a couple of weeks ago, like grant-building agents. These tasks feel really human, but are really commoditized skills that everyone in nonprofits across the world has to do.

Many jobs are templates. Years ago, I would use Microsoft Word templates, like professional letterhead or thank-you note formats. Now, instead of these simple letters, we’re doing all-out quarterly financial reconciliations or grant proposals.

I thought it was pretty wild when I read this week that Harvey is using Claude for a lot of its agent harnessing. Harvey is used by over a thousand law firms and, I think, over a hundred thousand lawyers use it. I mean, if Harvey’s just a few steps away from being a Claude wrapper, it’s just nuts.

There’s plenty of money to be made in these interstitial moments where, if you’re an early adopter, you probably have 12 to 18 months of profit. But long term, I’m not sure any of these folks, including Harvey, are any different than the Garmin GPS getting eaten by CarPlay.

Also this week, Anthropic is rumored to be getting ready to launch Orbit, which is a proactive assistant that generates briefings using connectors to your email, calendar, Drive, Slack, GitHub…basically anything you hook it up to. It can look at all of your systems each morning and give you a really nice briefing.

Google has something similar for Gmail called cc. OpenAI has Pulse that’s also similar.

I like all of them, to be honest. It’s not that complicated. They doing what you would do: checking all your stuff and giving you a little summary. You can even give it feedback and fine-tune it.

Anthropic added some powerful features to its Claude Managed Agents. I’m starting to feel behind again because I need to start building agents with Claude, but I haven’t had time.

The two additions this week are worth knowing. One lets Claude go back and review what it’s done and kind of figure out what it’s been doing well or should improve. I don’t know why Anthropic felt the need to call this feature “dreaming,” but I guess it is kind of like humans, right? It’s a scheduled process, just like sleep would be.

With “dreaming” every now and then, Claude goes back and reviews all of your different agent sessions and looks for patterns. Then it tries to figure out what it likes about what it did, when it made mistakes, whether those mistakes happened more than once, what it can do to fix them, what kind of bottlenecks occurred, and basically tries to better itself, for lack of a better term.

Self-reflection is what I would have called it.

The other feature is called “outcomes,” which is better named. That’s where you write a rubric saying what you think success looks like for your agent, and the agent will not stop working until it determines that it’s nailed what you said.

If I want to run a marathon, that’s 26.2 miles. But if I want to run a marathon under four hours, that’s a more specific goal. That’s measurable.

So the idea of telling my agent all of my different success metrics explicitly, and making sure that before the agent calls it a day, the agent hits all those metrics, is pretty valuable.

And if you combine that with dreaming or self-reflection, it goes back and sees how it could do better when it fails. That’s a pretty virtuous cycle.

It’s when I was reading this press release, where I saw that Harvey was using Claude managed agents. Also, Netflix is using them.

One of the battlegrounds for AI growth right now is business adoption, also known as enterprise use. Anthropic has started to corner the market over the past six months, and OpenAI has been trying to catch up. Google has sort of been, in my impression, waiting it out.

This week, even though they are ahead, Anthropic announced a $1.5 billion consulting venture to drive enterprise AI services. The launch partners are big enterprise partners, like Blackstone, Hellman & Friedman, and Goldman Sachs.

This AI services venture’s job is to basically collude, for lack of a better term, to try to get mid-sized companies to use Claude. So if you hire Goldman Sachs to do some consulting, they’re going to tell you how you can get Claude into your mid-sized partnership.

OpenAI is doing the same kind of thing, but with a $10 billion venture. We’ll get to that in a bit.

Closing out the major stories from Anthropic is the fact that Claude can now operate your Mac to complete desktop tasks. Computer use is nothing completely new, but it’s getting better and better… and now you can call or text your laptop to get it to work for you.

In this case, Claude can open up your apps, navigate your browsers, fill in spreadsheets, and do stuff you would do at your desk, but you’re texting Claude from your phone. When you get to your computer, it’ll be all done.

Shifting gears, I want to shine a light on Google’s Gemma.

Google Gemma is an incredibly powerful small model that can run locally and is absolutely crushing benchmarks. If you can embed Gemma into your app and use it as a little service running locally without a connection to the internet, you don’t have to pay token fees.

So I would imagine, if you’re an Android user, over time, Google is just going to have this engine on your Android, and then all these apps can basically have a free large language model that they can use that’s good enough for taking verbal commands.

Because a lot of language model skills are just taking language and turning it into action. It doesn’t have to be writing War and Peace. It could just understand that you want to open up Spotify.

So if you want quick speech-to-text understanding and natural language processing, Google Gemma is pretty incredible. And what’s even crazier is that it’s multimodal, so it can look at and understand what’s in a picture. So you could say, “I want you to circle the cat in this picture,” or “I want you to remove the person in the background of this picture.”

Basic editing and retouching can all be conversationally empowered by Gemma as well.

So I think what we’re going to see is a lot of agents are going to start using Gemma as their pass-through first filter. And then, of course, if they need to bring in the big guns, they can hit an API over the internet and start using its tokens.

But this combination of locally hosted, powerful models with moderated use of cloud-based models is only going to continue. And Google is just pioneering this right now.

In a world of non-US open local models like variants of DeepSeek, Qwen, Zai, and Mistral, Gemma is holding its own.

And this week, Google tripled Gemma 4’s inference speeds using something called multi-token prediction. All it really means is that Google has given Gemma little miniature helpers that do their best to guess as much as they can instantly, and then Gemma just confirms their work.

If you had to think about every single word you said with the same amount of effort, it would really slow you down because certain phrases are just strings that always go together. So instead of having Gemma’s engine think of each token one at a time, Goole has these multi-token prediction drafters that work as its intern, speeding up predictions and blurting out what they think is going to come next as fast as they can. 3x speed increase with no loss of quality.

The second big story from Google is that they’re rumored to be launching a unified video model that does both video and imagery. Right now Google uses Veo for video and Nano Banana for images.

What I don’t know is if this is just a convenient single-stop shop, or if it’s a new type of multimodality where the diffusion engine for images also diffuses video. If it’s the latter, that seems like a pretty big deal.

As we’ve seen in the past, there’s quite a bit of emerging intelligence coming from image and video tools. In the case of images, images can answer math questions as part of the diffusion, and image-to-image is so powerful that you can take a Google Maps screenshot, draw an arrow on it, and say, “What does the arrow see?” Then the image diffusion model will show you the view from the arrow’s vantage point!

When you combine that type of emergence into the multiple-vectors of video, it turns into physics understanding and world models. I’m not sure what this Omni model implies, but I’m going to keep an eye on it.

We talked a bit about Harvey. Harvey announced that they’re open-sourcing a long-horizon legal agent benchmark (LAB).

This benchmark will be a partnership between 15 frontier companies, all the way from NVIDIA to Google to Anthropic to LangChain to Mistral, as they try to understand the range of legal work that agents can accomplish.

The legal agent benchmark reminds me a little bit of the GDPval benchmark that OpenAI made for real-world tasks. However, this one is laser-focused on legal tasks.

LAB has over 1,200 agent tasks across 24 legal practice areas, with over 75,000 human-written rubric criteria.

Meta is partnering with a company called Overview Energy to bring up to one gigawatt of space solar energy back to Earth.

The Overview Energy company will put satellites in geosynchronous orbit about 22,000 miles above the Earth’s equator, and then beam the energy back to the ground as low-intensity near-infrared light. They aim it at solar farms that can grab the light and use it as energy.

It sounds to me like a directed beam, and I don’t know much about low-intensity infrared light, but it sounds like it’s not going to be a really bright spotlight, which is encouraging.

In addition, Meta is partnering with a battery company called Noon Energy that can store this energy coming from space.

Meta has acquired a humanoid robotics startup for its superintelligence lab. The lab has been running for about a year, and they just came out with their first frontier model, called MuseSpark.

Meta plans to kick the tires on Muse for a little bit, and then it sounds like they’re going to open source it, once it’s stable.

With the pace of Meta’s superintelligence lab quickening, it seems like they are going to get into the robotics world, as multimodal AI comes together as locally hosted robotic brains.

While we’re on the topic of robots, Figure posted a demonstration where its 03 robot walked up and down stairs using nothing but its onboard camera.

That’s an interesting counterpoint to one of the Figure demos from a few months ago, where they had the robot walk blindfolded through a giant debris field of broken-up pallets, sand, and uneven surfaces.

So I guess if you combine being able to walk blindfolded with being able to walk using only the onboard cameras, the Figure robot brain, HELIX, is becoming pretty robust.

Switching gears, over the past few weeks, Anthropic has made a stand against some of the requests from the Pentagon, and the Pentagon has been punitive in calling Anthropic a supply-chain risk for not complying.

That said, Anthropic has the best models and the momentum right now. They’ve had 80x annualized revenue growth when they were only predicting 10x growth, and they’re basically the leading frontier company, despite being labeled a supply-chain risk. It’s like watching Survivor.

This week the Pentagon reached agreements to deploy artificial intelligence on classified networks with eight technology companies. Those eight are Amazon Web Services, Google, Microsoft, NVIDIA, OpenAI, Oracle, Reflection, and SpaceX.

Moving back to the frontier labs, OpenAI, not to be left out, has done quite a bit this week as well.

We talked about Anthropic’s enterprise play with a bunch of venture backing, and it turns out OpenAI did the same thing…this same week. OpenAI launched a $10 billion AI venture backed by TPG, Bain, and SoftBank.

I’m pretty sure the idea is similar to what Anthropic is doing. By partnering with all these folks, they’re leveraging alliances to encourage businesses to adopt enterprise accounts with OpenAI. The more that businesses start working directly with OpenAI or Anthropic, the less that the wrapper companies are going to be able to keep up.

OpenAI also announced that ChatGPT can plug directly into Excel and Google Sheets, and Codex can build Google Slides live without opening your browser. These direct integrations of multimodal office features (with Google’s and Microsoft’s support) are pretty amazing.

Lastly, in OpenAI news, OpenAI launched three real-time voice models that include GPT-5 reasoning.

Real-time means real time (!)… which means the audio latency is low enough that you feel like you’re having a real conversation. So these are models that can reason, translate, and transcribe in real time.

The three models are: GPT Real-Time 2, which provides naturally paced conversations that have harder requests built in, so the thinking is so fast you don’t know it thought. Look out Harry Mack!

Then there’s the translation model, which is focused directly on real-time output. This supports 70 input languages and 13 outputs that can keep pace with natural conversation!!

And then there’s GPT Real-Time Whisper, which is the real-time speech-to-text model that transcribes as fast as you can speak.

This type of agent enablement is going take people’s jobs. If you just need to complete a transaction with a business where someone’s looking up information or entering information, there’s a good chance that an AI can look up that information just as fast, with a more chipper attitude, and never gets tired, and works around the clock.

The idea of business hours might go away. That is going to be a paradigm shift. What if we can just call our doctor at 2 in the morning and change an appointment, with as good of an experience as if we spoke to a human?

This week Perplexity has integrated some strong finance agent tools with officially licensed market data from partners like Morningstar and PitchBook.

That’s it for the largest stories. There are still over a dozen more that are worth knowing, below. Have a wonderful week!

Anthropic

Lead
Dario Amodei: I think the other major U.S. AI labs are roughly 1–3 months behind Anthropic in terms of capabilities. Dario Amodei: There is no issue with putting more compute into Mythos. Dario Amodei: Anthropic’s revenue grew 80x on an annualized basis as of Q1.
https://x.com/jukan05/status/2051847480254570998?s=12

Growth
Dario Amodei: We planned for a world of 10X growth per year. In the first quarter of this year, we saw .. 80X (annualized) growth per year in revenue and usage. That is the reason we had difficulties with compute
https://x.com/firstadopter/status/2052118224888607107

Dario&Daniela Amodei Interview: The 80x growth reportedly caught them completely off guard, and that’s the reason for the compute constraint. The compute deal with SpaceX is the first attempt to address the shortage and continues to search for solutions.
https://x.com/kimmonismus/status/2052118418174681572

Google Chips
Anthropic Commits to Spending $200 Billion on Google’s Cloud and Chips — The Information
https://www.theinformation.com/articles/anthropic-commits-spending-200-billion-googles-cloud-chips

Colossus
Anthropic is getting 220,000 Nvidia GPUs from SpaceX and has committed to spending $200 Billion on Google TPUs. Claude Code’s 5-hour rate limits are doubling. Dario is finally getting aggressive about chips from every possible source to fight OpenAI Codex’s growth.
https://x.com/Yuchenj_UW/status/2052065017072386450

Anthropic just announced a new partnership with SpaceX for additional compute. The company is upping usage limits by 2x for Claude Code across plans, removing peak hours reductions, and raising API limits for Opus.
https://x.com/TheRundownAI/status/2052064469371470218

Anthropic rents out the entire Colossus 1 datacenter from xAI / SpaceX with over 220,000 Nvidia GPUs including approximately 150,000 H100 GPUs, 50,000 H200 GPUs, and 30,000 GB200
https://x.com/scaling01/status/2052068218047545501

Elon’s making $500 billion off Anthropic SpaceX numbers on this ? $18.5 billion revenue gunning for $1.75 trillion IPO valuation 95x Price/Sales ratio Renting the GPUs to Anthropic ~ $5 billion in additional revenue x 95 = $475 billion in valuation credit for the IPO
https://x.com/andrewbenson/status/2052147078902718583?s=46

Higher usage limits for Claude and a compute deal with SpaceX \ Anthropic
https://www.anthropic.com/news/higher-limits-spacex

In the next few days we’ll be ramping up Claude inference on Colossus. Grateful to be partnering with SpaceX here. We are going to need to move a lot of atoms in order to keep up with AI demand, and there’s nobody better at quickly moving atoms (on or off planet Earth)
https://x.com/nottombrown/status/2052062566126649448

not even 3 months ago Elon was rabidly chudposting at Amanda Askell now, all water under the bridge! (water goes into Colossus to be annihilated by Anthropic’s inference demand) Not evil at all! Highly impressive people!
https://x.com/teortaxesTex/status/2052080900280557749

SpaceXAI will provide @AnthropicAI with access to Colossus 1, one of the world’s largest and fastest-deployed AI supercomputers, to provide additional capacity for Claude →
https://x.com/xai/status/2052060350770515978

Under-reported details of the xAI/Anthropic Colossus data center deal: Anthropic get Colossus 1 but xAI keep using the larger Colossus 2, Colossus 1 has a REALLY bad environmental record, and xAI just shut down a bunch of older models on 2 weeks’ notice
https://x.com/simonw/status/2052436629365948920

Very excited to be partnering with @elonmusk @SpaceX! Visionary engineering + Claude is going to be awesome, scaling is continuing for a long time!
https://x.com/Mononofu/status/2052212359536496961

We’ve agreed to a partnership with @SpaceX that will substantially increase our compute capacity. This, along with our other recent compute deals, means that we’ve been able to increase our usage limits for Claude Code and the Claude API.
https://x.com/claudeai/status/2052060691893227611

went under the radar with the spaceX <> anthropic compute deal, but those features make a ton of sense for power users and i’d expect them to become very important in the near future dreaming: better and more expressive memory i’d expect, with cross session context do detect
https://x.com/eliebakouch/status/2052156107313807690

Finance Agents
Agents for financial services \ Anthropic
https://www.anthropic.com/news/finance-agents

New agents for financial services | Claude Cowork + Claude Managed Agents – YouTube
https://www.youtube.com/watch?v=foxeK2AXfHQ

There goes another bunch of startups: Anthropic launched pre-built agent templates for financial services that handle tasks like valuation analysis, KYC screening, and month-end close, packaged with connectors to major data providers like FactSet, S&P Global, and Morningstar.
https://x.com/kimmonismus/status/2051681279582540114

Orbit
Anthropic Orbit leaked Orbit, a proactive assistant for Claude Cowork that auto-generates briefings and insights from Gmail, Slack, GitHub, Calendar, Drive, and Figma, no prompting required. Users can also deploy and pin “”Orbit apps”” for quick access. It’s Anthropic’s answer to
https://x.com/kimmonismus/status/2051618156385366305

Anthropic working on Orbit, its upcoming proactive assistant
https://www.testingcatalog.com/anthropic-is-working-on-orbit-its-upcoming-proactive-assistant/

Agents Platform
Anthropic is building out their managed agents platform, adding Dreaming (memory) and Outcomes (rubrics). The idea I’m wrestling with: how differentiated are these platform features really? I initially thought the model would “”eat your scaffolding””, an argument best made by
https://x.com/RichNwan/status/2052085746526216601

New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration | Claude
https://claude.com/blog/new-in-claude-managed-agents

Financial Services
Anthropic Unveils $1.5 Billion Joint Venture With Wall Street Firms – WSJ
https://www.wsj.com/business/deals/anthropic-nears-1-5-billion-joint-venture-with-wall-street-firms-8f5448ee

Building a new enterprise AI services company with Blackstone, Hellman & Friedman, and Goldman Sachs \ Anthropic
https://www.anthropic.com/news/enterprise-ai-services-company

Google in talks with Blackstone, KKR, EQT for omnibus Gemini AI licensing as OpenAI and Anthropic build consulting ventures
https://thenextweb.com/news/google-blackstone-kkr-omnibus-ai-licensing-private-equity

New for financial services: ready-to-run Claude agent templates for building pitches, conducting valuation reviews, closing the books at month-end, and more. Install them as plugins in Cowork and Claude Code, or use our cookbooks to run them in production as Managed Agents.
https://x.com/claudeai/status/2051679629488865498

The finance sector is Anthropic’s second highest revenue contributor by industry
https://x.com/madisonmills22/status/2051688936053813661?s=46

Computer Use
You can now enable Claude to use your computer to complete tasks. It opens your apps, navigates your browser, fills in spreadsheets—anything you’d do sitting at your desk. Research preview in Claude Cowork and Claude Code, macOS only.
https://x.com/claudeai/status/2036195789601374705?s=20

Cowork: Claude Code power for knowledge work | Claude by Anthropic
https://claude.com/product/cowork#dispatch-and-computer-use

Google

Gemma
Gemma 4 just got even faster! We’re releasing Multi-Token Prediction (MTP) drafters that deliver up to a 3x speedup, without any degradation in output quality or reasoning logic.
https://x.com/googlegemma/status/2051713412431007808

Gemma 4 up to 3x faster, directly in your phone! 🚀 Check out the difference Speculative Decoding makes! Multi-Token Prediction (MTP) is supercharging inference speeds for Gemma 4.
https://x.com/googlegemma/status/2052468624657654194

Gemma 4: Now up to 3x Faster. ⚡ Same quality, way more speed. Our new MTP drafters allow Gemma 4 to predict multiple tokens at once, effectively tripling your output speed without compromising intelligence.
https://x.com/googledevs/status/2051700498328346945

Gemma-4 lands in Code Arena: Frontend Webdev and shifts the Pareto Frontier! Among open models, Gemma-4-31b ranks #13 and Gemma-4-26b-a4b ranks #17. Congrats to @GoogleDeepMind on shifting the frontier!
https://x.com/arena/status/2052061349312921686

I benchmarked Google’s new MTP for Gemma 4 31B using vLLM with 4 speculative tokens, a fairly conservative setup. Results: – Much higher throughput than Qwen3.6’s MTP – Lower latency too, helped by Gemma 4 generating fewer tokens – For coding tasks with reasoning enabled,
https://x.com/bnjmn_marie/status/2052286398707687650

Multi-token-prediction in Gemma 4
https://blog.google/innovation-and-ai/technology/developers-tools/multi-token-prediction-gemma-4/

Omni
Google is testing new Omni model for video generation
https://www.testingcatalog.com/google-is-testing-new-omni-model-for-video-generation-ahead-of-i-o/

Benchmark

Harvey
Artificial Analysis is partnering with Harvey on their new Legal Agent Benchmark! Harvey’s Legal Agent Benchmark (LAB) is an agent-native take on how AI should be contributing to legal work in 2026 – made up 1200 agentic tasks across 24 practice areas. It’s highly aligned with
https://x.com/ArtificialAnlys/status/2052145762650431840

Introducing Harvey’s Legal Agent Benchmark
https://www.harvey.ai/blog/introducing-harveys-legal-agent-benchmark

LAB is the first long-horizon, open-source legal agent benchmark, from @harvey. it will help legal teams answer “”what can legal agents do today?””, plan deployment, and design human-agent cooperation. autonomous legal is a deep domain, and a good benchmark can accelerate progress
https://x.com/saranormous/status/2052061665596948894

Meta

Data Centers
Meta is planning to power its AI data centers with solar energy beamed from space. If it works, solar farms could produce power 24/7 without batteries or backup generators. The company behind it all is Overview Energy — they want to launch 1,000 satellites into orbit, 22,000
https://x.com/rowancheung/status/2051320518905930208

OpenAI

Agent Phone
https://about.fb.com/news/2026/04/powering-ai-strengthening-the-grid-space-solar-energy-and-long-duration-storage/

Figure

Robot
F.03 can now walk up/down stairs purely using it’s onboard camera perception Our robots now walk from manufacturing when built to HQ This is trained end-to-end with reinforcement learning in simulation
https://x.com/adcock_brett/status/2050624857730417097

Government

Pentagon
Meta-Backed Scale AI Wins $500 Million Defense Department Deal – Bloomberg
https://www.bloomberg.com/news/articles/2026-05-06/meta-backed-scale-ai-wins-500-million-defense-department-deal

Pentagon strikes AI deals for classified military use – The Washington Post
https://www.washingtonpost.com/technology/2026/05/01/pentagon-ai-deals-microsoft-amazon-google-classified-military/

Today, the @DeptofWar entered into agreements with SEVEN of the world’s leading frontier AI model and infrastructure companies to deploy frontier capabilities on the Department’s classified networks: • SpaceX • OpenAI • Google • NVIDIA • Reflection • Microsoft • Amazon
https://x.com/DoWCTO/status/2050175912134561977

Meta

Robots
Meta buys robotics startup to bolster its humanoid AI ambitions | TechCrunch
https://techcrunch.com/2026/05/01/meta-buys-robotics-startup-to-bolster-its-humanoid-ai-ambitions/

Meta is making a serious push into AI for humanoid robots Meta has acquired ARI, a robotics AI startup, to accelerate its humanoid technology efforts. The deal closed today; financial terms were not disclosed. The ARI team (including co-founders and staff from San Diego and New
https://x.com/TheHumanoidHub/status/2050340597685518461

OpenAI

Excel and Sheets
ChatGPT is now available as an add-on in Excel and Google Sheets. It can help analyze messy data, write formulas, update spreadsheets, and explain what it’s doing along the way—without leaving your spreadsheet. Powered by GPT-5.5.
https://x.com/ChatGPTapp/status/2051776032127238266

Create Google Slides in Codex without opening your browser, clicking buttons, and manually aligning figures. Plus, you (and your team) can view the progress in realtime. Codex isn’t creating the deck locally, then uploading it. It’s actually iteratively building it, checking
https://x.com/gabrielchua/status/2051113129317408925

Venture
OpenAI launches $10B AI venture backed by TPG, Bain, SoftBank – Bloomberg
https://www.msn.com/en-us/money/general/openai-launches-10b-ai-venture-backed-by-tpg-bain-softbank-bloomberg/ar-AA22miSj

RealTime

🎙️ Voice AI only feels natural when conversation keeps pace with speech. Here’s how we rebuilt our WebRTC stack with a thin relay and stateful transceiver to keep real-time media fast for ChatGPT voice, the Realtime API, and more.
https://x.com/OpenAIDevs/status/2051453905343828350

🚀 GPT-Realtime-2 just landed in Genspark. Our Call for Me Agent now runs on it. Genspark Realtime Voice is upgrading next. What Realtime 2 brings: Sharper reasoning. Tighter instruction following. +26% effective conversation rate. Far fewer dropped calls.
https://x.com/genspark_ai/status/2052524670088556557

Advancing voice intelligence with new models in the API | OpenAI
https://openai.com/index/advancing-voice-intelligence-with-new-models-in-the-api/

Building voice applications with GPT-Realtime-2? Our new prompting guide covers how to tune reasoning effort, use preambles, design tool behavior, handle unclear audio, capture exact entities, and maintain state in longer sessions.
https://x.com/OpenAIDevs/status/2052530378184032560

Congrats to @OpenAI for taking the top spot on our Audio MultiChallenge S2S leaderboard with the release of GPT‑Realtime‑2 🥇 GPT-Realtime-2 more than doubles GPT-Realtime-1.5 on instruction retention, rising from 36.7% to 70.8% APR, and also stands out on voice editing,
https://x.com/ScaleAILabs/status/2052451341071683732

Dubbing for live events… in real time? 😮 Here’s OpenAI’s new GPT-Realtime-Translate model in action in Vimeo. Those translations are happening completely live. No pre-loaded captions. Live dubbing is one of the many features we’re exploring this year… (Hopefully) more
https://x.com/Vimeo/status/2052442588201029684

GPT-Realtime-2 audio input price remains steady at $1.15 per hour of audio input, and $4.61 per hour of audio output.
https://x.com/ArtificialAnlys/status/2052486478501204415

gpt-realtime-2 shows a 15pp improvement (vs 1.5) on Big Bench Audio, and is now close to saturation.
https://x.com/juberti/status/2052507302092296252

GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper are available in the Realtime API today.
https://x.com/OpenAIDevs/status/2052440968763515223

GPT-Realtime-2: Building a Live Translator
https://x.com/RayFernando1337/status/2052479718495318143

GPT-Realtime-Whisper brings low-latency streaming transcription to the Realtime API. Use it when your app needs to understand speech continuously while the interaction is still unfolding.
https://x.com/OpenAIDevs/status/2052440957258489859

Guess who’s back, back again. Whisper, but now with realtime streaming. Check out the new gpt-realtime-whisper transcription model in my
https://t.co/b2UTuSxhOI demo.
https://x.com/juberti/status/2052478775523512356

have been excited for realtime voice-to-voice translation as an AI application since we started OpenAI. extremely cool to see it now available in the API for anyone to build with:
https://x.com/gdb/status/2052480998668206262

Introducing GPT-Realtime-2 in the API: our most intelligent voice model yet, bringing GPT-5-class reasoning to voice agents. Voice agents are now real-time collaborators that can listen, reason, and solve complex problems as conversations unfold. Now available in the API
https://x.com/OpenAI/status/2052438194625593804

New ChatGPT Voice mode pretty much confirmed. And im really excited for it.
https://x.com/kimmonismus/status/2051571219040735423

New Voice Model from OpenAI in the API gpt-realtime-2 Here a quick demo I built
https://x.com/diegocabezas01/status/2052492653082681485

OpenAI has released GPT-Realtime-2, achieving 96.6% in our Speech Reasoning benchmark, Big Bench Audio, and #1 in our Conversational Dynamics benchmark Released today, GPT-Realtime-2 is OpenAI’s new flagship native Speech to Speech model, introducing adjustable reasoning effort
https://x.com/ArtificialAnlys/status/2052486470469140777

OpenAI shipped a new speech-to-speech model today: gpt-realtime-2 This is the first speech-to-speech model good enough to use in my voice agents that do “”real work.”” Or real play, for that matter. Here’s gpt-realtime-2 as the brain of the ship AI in Gradient Bang. The
https://x.com/kwindla/status/2052521318688739811

Our new voice models are now available in the Realtime API: 🎙️ GPT-Realtime-2: Build production-ready voice agents that can think harder, take action, handle interruptions, and keep conversations flowing. 🎙️ GPT-Realtime-Translate: Translate while streaming across more than 70
https://x.com/OpenAI/status/2052438196454379986

people are really starting to use voice to interact with AI, especially when they have a lot of context to dump. GPT-Realtime-2 comes to the API today; it is a pretty big step forward. (we are working on improvements to voice in chat.)
https://x.com/sama/status/2052462271667028211

pretty excited for voice models to get great its interesting to watch how people are already starting to change the way they interface with AI
https://x.com/sama/status/2051464865634742334

Saw this and thought “”yes! ChatGPT voice mode is going to stop acting like a two-year-model”” but that upgrade hasn’t shipped just yet
https://x.com/simonw/status/2052439091577496054

Taking talking shop to a whole new level. We just shipped Glean’s real-time voice capability, powered by @OpenAI’s newest speech model GPT-Realtime-2. Grounded in the context across your org, it feels like a real AI coworker and can keep up with how work gets finished. In
https://x.com/glean/status/2052440702169108990

Updated my hello-realtime demo to use the new gpt-realtime-2 model (now with reasoning). Check it out at
https://t.co/td6Cx2EOPO, or call 425-800-0042!
https://x.com/juberti/status/2052469176821002676

Using @OpenAI gpt-realtime-2 to get a glimpse of future voice-first experiences. A market dashboard you don’t click through. You direct it. Say, “Focus on Apple,” and the whole interface changes. Ask, “How did it do over the last 30 days?” and the chart updates. Say, “Go
https://x.com/levinstanley/status/2052506605044842672

Voice agents are getting more capable. Here’s what’s new: • GPT-Realtime-2 for voice agents that reason and take action • GPT-Realtime-Translate enabling translation from 70 input languages into 13 output languages • GPT-Realtime-Whisper, making transcription even faster
https://x.com/OpenAIDevs/status/2052440907933474954

Voice agents are so back!! Today we’re launching 3 new realtime audio models in the API: 🎙️ GPT-Realtime-2 GPT-5-class reasoning for voice agents that can use tools, recover from interruptions, and carry longer conversations with 128K context 🌍 GPT-Realtime-Translate Live
https://x.com/reach_vb/status/2052438371058737280

Voice workflows just got stronger with gpt-realtime-1.5 in the Realtime API. The model offers more reliable instruction following, tool calling, and multilingual accuracy. Demo with @charlierguo
https://x.com/OpenAIDevs/status/2026014334787461508

We know you’re eager for voice updates in ChatGPT. Stay tuned, we’re cooking.
https://x.com/OpenAI/status/2052438197695877316

Perplexity

Finance
Finance Search is now available in the Perplexity Agent API. In one tool call, developers can now retrieve licensed financial datasets, real-time market data, and cited web sources for agents that need current, verifiable financial answers.
https://x.com/perplexity_ai/status/2052028012313649194

Perplexity Computer now brings in licensed data for professional finance research and analysis with 35 dedicated workflows that are representative of the daily work of an analyst
https://x.com/AravSrinivas/status/2051694381137350661

Today we’re launching Perplexity Computer for Professional Finance. Finance teams can bring licensed data from providers like Morningstar, PitchBook, Daloopa, and Carbon Arc into Computer. We’ve also added 35 dedicated finance workflows for the work analysts repeat every week.
https://x.com/perplexity_ai/status/2051693893473935372

Alignment

Harness
we’re continuing to see clear examples where a model’s harness is a major determinant of overall performance. with the same model, running on same task, it’s easy to observe very different scores depending on (system) prompts, tools (& their descriptions), and middleware
https://x.com/masondrxy/status/2052054177749029164

AllenAI

Robots
MolmoAct 2: An open foundation for robots that work in the real world | Ai2
https://allenai.org/blog/molmoact2

Real-world robots that actually work in messy kitchens, labs, or factories… fully open-source and ready to deploy today: A model that can do toy demos? Nope. AI2 just dropped MolmoAct 2: a bimanual action reasoning model that handles real chores (washing dishes, sorting items)
https://x.com/IlirAliu_/status/2051934034935128446

Anthropic

Enterprise
Anthropic and OpenAI are both launching joint ventures for enterprise AI services | TechCrunch
https://techcrunch.com/2026/05/04/anthropic-and-openai-are-both-launching-joint-ventures-for-enterprise-ai-services/

Both Anthropic and OpenAI have new initiatives to help enterprises deploy AI agents within their organizations. This is a trend that’s early but going to get very big fast. As agents enter knowledge work beyond coding, there is very real work to upgrade IT systems, get agents
https://x.com/levie/status/2051344780328858040?s=46

Institute
Focus areas for The Anthropic Institute \ Anthropic
https://www.anthropic.com/research/anthropic-institute-agenda

We’re sharing the research agenda of The Anthropic Institute, or TAI. TAI will focus on four areas: 1) Economic diffusion 2) Threats and resilience 3) AI systems in the wild 4) AI-driven R&D Read the full agenda:
https://x.com/AnthropicAI/status/2052385812881228218

Jupiter
Anthropic tests Jupiter-v1-p before potential launch in May
https://www.testingcatalog.com/anthropic-tests-jupiter-v1-p-before-potential-launch-on-may-6/

Mythos
With the help of Claude Mythos Preview, the Firefox team fixed more security bugs in April than in the past 15 months combined.
https://x.com/alexalbert__/status/2052468573516513762

Business Moats
A critical question in agent design is “how do we build agentic workflows so humans are given significant, interesting, or variance-producing decisions as they come up in the work?” A Claude-run company has no source of competitive advantage compared to other Claude-run firms.
https://x.com/emollick/status/2052066205226123472

Colossus Impact
I usually avoid commenting too much on industry deals, but this one is fascinating. Certainly seems like a blow to the idea that Grok will remain a frontier model.
https://x.com/emollick/status/2052068290311139769

Apple

Lawsuit
Apple to pay $250m to iPhone buyers over AI features lawsuit
https://www.bbc.com/news/articles/c0j2nydnzy7o

iOS27
Apple plans to make iOS 27 a Choose Your Own Adventure of AI models | TechCrunch
https://techcrunch.com/2026/05/05/apple-plans-to-make-ios-27-a-choose-your-own-adventure-of-ai-models/

Benchmarks

Refactoring
Today we’re releasing Refactoring, the final leaderboard of our SWE Atlas suite. This new leaderboard is the ultimate test of an agent’s ability to restructure code without breaking the system. Claude Opus 4.7 with Claude Code takes the top spot🥇
https://x.com/ScaleAILabs/status/2052434456510878021

Business

Coinbase
This is an email I sent earlier today to all employees at Coinbase: Team, Today I’ve made the difficult decision to reduce the size of Coinbase by ~14%. I want to walk you through why we’re doing this now, what it means for those affected, and how this positions us for the
https://x.com/brian_armstrong/status/2051616759145185723?s=20

Doctors
Missing from the “will AI replace doctors?” debate is that doctors (and lawyers and psychologists and bankers) all vote & form the donor base to political parties & have deep community ties. The government will largely determine what AI is allowed to do, no matter what it can do
https://x.com/emollick/status/2051684693838340470

Sierra
Sierra is raising $950 million from new and existing investors, led by Tiger Global and GV, at a valuation of over $15 billion. We now have more than $1 billion to invest in becoming the global standard for companies wanting to transform their customer experiences with AI.
https://x.com/btaylor/status/2051313954312331411

Computing

China Smuggling
How much AI compute has been smuggled to China? We estimate between 290k and 1.6M H100-equivalents by the end of 2025 — representing ~20% to ~60% of China’s total compute.
https://x.com/EpochAIResearch/status/2049924785153638761

Google

Film Prize
Submissions are open for the Future Vision XPRIZE
https://blog.google/innovation-and-ai/technology/ai/future-vision-film-competition-xprize/

EVE
Google DeepMind partners with EVE Online for AI model testing – Ars Technica
https://arstechnica.com/gaming/2026/05/google-deepmind-partners-with-eve-online-for-ai-model-testing/

Studio Behind EVE Online Goes Independent, Rebrands as Fenris Creations, Enters Research Partnership with Google DeepMind – Fenris Creations
https://www.ccpgames.com/news/2026/studio-behind-eve-online-goes-independent-rebrands-as-fenris-creations-enters-research-partnership-with-google-deepmind

AlphaEvolve
AlphaEvolve: Gemini-powered coding agent scaling impact across fields — Google DeepMind
https://deepmind.google/blog/alphaevolve-impact/

Gemini Flash Lite
gemini 3.1 flash-lite is here it’s our most cost-efficient model, optimized for high-volume agentic tasks, translation, and simple data processing
https://x.com/GoogleAIStudio/status/2052453828272812310

Health
📣 We’re continuing to bring more value to our AI Pro and Ultra subscribers! Google Health Coach is unlocked by Google Health Premium, which is going to be included at no extra cost in our Google AI Pro and Ultra plans starting on May 26. I’ve been dogfooding the Fitbit Air and
https://x.com/shimritby/status/2052439569136767291

Get up close and personal with your health. On May 26, the Fitbit app becomes the #GoogleHealth app for both Android and iOS— combining the best of Fitbit tracking with the power of Google to create a more holistic wellness experience. Learn more:
https://x.com/googlehealth/status/2052392762255761701

Introducing Fitbit Air. It’s lightweight, screenless and comfortable enough to wear 24/7 — with a battery life* of up to a week. * Battery life depends upon many factors and usage and actual battery life may be lower.
https://x.com/Google/status/2052501704155775481

Introducing the Google Health app
https://blog.google/products-and-platforms/products/google-health/google-health-app/

Workplace CLI
gog 0.16 is out. Google Workspace CLI for humans and agents. Lossless raw API output, sanitized Gmail reads, safer command profiles, Drive inventory, Docs tabs, Sheets tables, Gmail filter export, and official Docker images.
https://x.com/steipete/status/2051575048348074450

Antigravity
Google tests screen sharing and custom agents in Antigravity
https://www.testingcatalog.com/google-tests-screen-sharing-and-custom-agents-in-antigravity-ide/

File Search
Gemini API File Search is now multimodal
https://blog.google/innovation-and-ai/technology/developers-tools/expanded-gemini-api-file-search-multimodal-rag/

Good news for AI builders: the File Search tool in the Gemini API is now multi-modal 🗃️, powered by our Gemini Embedding 2 model, + support for custom metadata & inline citations : ) File Search comes with storage and embedding generation at query time free of charge!
https://x.com/OfficialLoganK/status/2051728186824904743

The Gemini API’s File Search tool now supports multimodal retrieval. Use `gemini-embedding-2` as the embedding model to build a true multimodal RAG system for PDFs and images with a single call. How it works: 1. Create a store with `gemini-embedding-2` as the embedding model 2.
https://x.com/_philschmid/status/2052060912425546050

Government

Maryland
Maryland Is First to Ban A.I.-Driven Price Increases in Grocery Stores – The New York Times
https://www.nytimes.com/2026/05/01/business/surveillance-pricing-groceries-maryland.html

Vettting
White House Considers Vetting A.I. Models Before They Are Released – The New York Times
https://www.nytimes.com/2026/05/04/technology/trump-ai-models.html

Meta

Agent
Meta prepares Hatch AI Agent with waitlist and social skills
https://www.testingcatalog.com/meta-prepares-hatch-agent-under-waitlist-and-social-media-skills/

OpenAI

Codex
Codex has surpassed Claude Code in downloads. According to TickerTrends, the crossover happened on April 30, after which Codex continued to gain share while Claude Code’s growth visibly slowed. Claude 4.7 was released April 16th, GPT-5.5 April 24th. Connect the dots.
https://x.com/kimmonismus/status/2051515496567292310

I’ve never used an agent for the cliches of ordering food, grocery shopping, or booking travel. But I repeatedly use Computer Use in Codex to add things to my family calendar in Apple Calendar. Like, I gave it my son’s little league schedule for the next four months, and it
https://x.com/_simonsmith/status/2050178967735353837

One week since the launch of GPT-5.5, and it’s already our strongest model launch yet. API revenue is growing more than 2x faster than any prior release, while Codex doubled revenue in under seven days as enterprise demand for agentic coding tools keeps climbing.
https://x.com/OpenAI/status/2050250926888468929

Elon
Mira Murati tells the court that she couldn’t trust Sam Altman’s words | The Verge
https://www.theverge.com/ai-artificial-intelligence/925338/openai-musk-v-altman-mira-murati

Musk sought settlement with OpenAI two days before trial
https://www.cnbc.com/2026/05/04/musk-altman-open-ai-settlement-trial-brockman.html

MRC
MRC is already deployed across all of OpenAI’s largest supercomputers that we use to train frontier models, including our site with @Oracle Cloud Infrastructure (OCI) in Abilene, Texas, and in @Microsoft’s Fairwater supercomputers. MRC is now available through the
https://x.com/OpenAI/status/2052025533937103102

NVIDIA just open-sourced a transport protocol that powers OpenAI’s Blackwell clusters. It opened MRC, a new RDMA transport protocol for massive AI training clusters. Instead of pushing GPU traffic through one fragile path, MRC spreads a single connection across multiple network
https://x.com/kimmonismus/status/2052011784023028060

Supercomputer networking to accelerate large scale AI training | OpenAI
https://openai.com/index/mrc-supercomputer-networking/

We’ve partnered with @AMD, @Broadcom, @Intel, @Microsoft, and @NVIDIA, to release Multipath Reliable Connection (MRC), a new open networking protocol that helps large AI training clusters run faster and more reliably, with less wasted GPU time.
https://x.com/OpenAI/status/2052025532485902368

Nanny
Introducing Trusted Contact in ChatGPT | OpenAI
https://openai.com/index/introducing-trusted-contact-in-chatgpt/

Enterprise iOS app
OpenAI releases a separate ChatGPT iOS app for enterprise users – 9to5Mac
https://9to5mac.com/2026/05/04/openai-releases-a-separate-chatgpt-ios-app-for-schools-and-work-organizations/

Instant
5.5 instant comes to ChatGPT today! imo it is a pretty big upgrade, i really like using it.
https://x.com/sama/status/2051716909629153573

All benchmarks are flawed, but GPQA has been fairly consistent & highly correlated with other measured benchmars. I think it’s a good way to see how far we’ve come that the free model from OpenAI, GPT 5.5 Instant, is at a level that even paid models did not reach until late 2025
https://x.com/emollick/status/2051801703209742734

Excited that we’re updating the default model in ChatGPT today! 5.5 instant is a substantial improvement in intelligence, image perception, and factuality. It also updates the writing style to be a bit plainer and more straightforward. What was on your wishlist?
https://x.com/ericmitchellai/status/2051711459886059963

GPT-5.5 Instant is rolling out over the next two days as the default model to all ChatGPT users, and as ‘gpt-5.5-chat-latest’ in the API. Personalization improvements are rolling out to Plus and Pro users on the web, and soon on mobile. Memory sources are rolling out across all
https://x.com/OpenAI/status/2051709035347694047

GPT-5.5 Instant is starting to roll out in ChatGPT. It’s a big upgrade, giving you smarter, clearer, and more personalized answers in a warmer, more natural tone. And it’s also more concise, which we heard you wanted. We think you’ll love chatting with it.
https://x.com/OpenAI/status/2051709028250915275

GPT-5.5 Instant: smarter, clearer, and more personalized | OpenAI
https://openai.com/index/gpt-5-5-instant/

the new instant model in chatgpt is so good damn if you have been thinking-model-only for awhile, give it a try!
https://x.com/sama/status/2051758152224506203

Agents SDK
Agents SDK 2.0 is underrated
https://x.com/sama/status/2050998576671859003

OpenAI Agents SDK – an open orchestration layer for building multi-agent workflows It lets you define agents as LLMs with instructions, tools (APIs, functions, external systems), guardrails, and supports: • sessions with conversation history management • human-in-the-loop •
https://x.com/TheTuringPost/status/2050903494010499113

Images
GPT-imagegen-2: “”make 5×5 grid of dog photos, where each photo gets noticeably cuter”” …now cats …now man-eating squid …now covers of the book the Great Gatsby
https://x.com/emollick/status/2050049582688538736

OpenClaw

Updates
Me and codex were busy. 🔊 https://t.co/FBNMbWOuFZ — Sonos 🗃️ https://t.co/YDdZyN2vwP — WhatsApp 🪶 https://t.co/eykEElx1Ez — X archive 🧰 https://t.co/txvYVtvhPg — GitHub archive 🛰️ https://t.co/2u2ACJEKKi — Discord archive 🎧 https://t.co/nrv2rzKfH4 — Spotify 💬
https://x.com/steipete/status/2051900143339704730

Perplexity

Medicine
Perplexity and Computer now allow you to run Deep and Wide Research on sources trusted by doctors and medical professionals like the New England Journal of Medicine, the British Medical Journal, the American Diabetes Association, and so on.
https://x.com/AravSrinivas/status/2051711236224761983

Perplexity and Computer now connect to premium health sources, starting with NEJM and BMJ Group, with 9 more medical journals and clinical databases on the way. Ask health questions and get answers cited from the same sources relied on by hospitals and research institutions.
https://x.com/perplexity_ai/status/2051710342242480538

Science

China Brain
China approves brain chip to overcome paralysis | Nature Biotechnology
https://www.nature.com/articles/s41587-026-03101-8

Standford Medicine
Medicine | The 2026 AI Index Report | Stanford HAI
https://hai.stanford.edu/ai-index/2026-ai-index-report/medicine

Twitter

Voices
Custom Voices and Voice Library | xAI
https://x.ai/news/grok-custom-voices

Try Grok Voice for your customer support
https://x.com/elonmusk/status/2052530063913189879

Two voices. One human. One AI. Can you guess the AI clone? 👇 Voice cloning, rich with natural emotion, is now live on the Grok Voice API.
https://x.com/xai/status/2051438210065322244

Grok 4.3
xAI has launched Grok 4.3, achieving 53 on the Artificial Analysis Intelligence Index with improved agentic performance, ~40% lower input price, and ~60% lower output price than Grok 4.20 The release of Grok 4.3 places @xAI just above Muse Spark and Claude Sonnet 4.6 on the
https://x.com/ArtificialAnlys/status/2049987001655714250

Full Executive Summaries with Links, Generated by Haiku 4.5

Anthropic claims significant lead over rivals with 80x revenue surge.
Anthropic’s CEO states the company is one to three months ahead of competitors in AI capabilities while signaling confidence in scaling up computing power for its next model, Mythos. The claim of a substantial capability gap is notable because it suggests one lab has pulled decisively ahead in the race to build more powerful systems, backed by an 80-fold revenue increase that enables continued investment.

Anthropic’s explosive growth outpaced its computing infrastructure by eightfold.
Anthropic CEO Dario Amodei revealed the company experienced 80X annualized growth in revenue and usage during early 2024—far exceeding the tenfold growth it had planned for—leaving it scrambling to secure enough computing power to meet demand. The unexpected surge prompted the company to pursue partnerships like a compute deal with SpaceX, illustrating how rapidly scaling AI services can create bottlenecks even for well-funded startups.

Claude maker Anthropic commits to major Google infrastructure spending.
Anthropic has pledged to spend $200 billion over multiple years on Google Cloud services and custom AI chips, cementing a deepening partnership that locks in the startup’s computing infrastructure for years. This unusually large commitment—rare for early-stage AI labs—suggests confidence in scaling Claude’s capabilities while reducing reliance on other cloud providers, though it also increases Anthropic’s financial obligations and dependence on Google’s roadmap.

Anthropic secures 220,000 Nvidia GPUs from SpaceX’s Colossus datacenter.
Anthropic announced a major compute partnership with SpaceX gaining access to over 220,000 GPUs—primarily high-end H100 and H200 chips—enabling the company to double Claude Code’s usage limits and significantly raise API rate limits for its most demanding customers. This deal is one of several major infrastructure investments (including commitments from Amazon, Google, and Microsoft) that underscore an intensifying arms race for AI compute capacity as companies race to scale their models and handle growing user demand.

Anthropic launches ten pre-built AI agents for financial services work.
Anthropic released ready-to-use AI agent templates designed to automate time-intensive financial tasks—including pitchbook creation, KYC compliance screening, and month-end accounting closes—that typically take months to build from scratch. The agents integrate directly into Microsoft Office tools (Excel, PowerPoint, Word, Outlook) and connect to major data providers like FactSet and S&P Capital IQ, allowing financial firms to deploy them in days rather than months. Major institutions including Citadel and FIS are already adopting Claude for these workflows, signaling potential displacement of custom-built automation and smaller fintech vendors serving this segment.

Anthropic launches Orbit, an auto-briefing assistant for Claude Cowork.
Orbit generates personalized daily insights from work tools like Gmail, Slack, GitHub, and Figma without requiring user prompts—positioning Anthropic’s offering as developer-focused compared to OpenAI’s ChatGPT Pulse, which launched similar proactive briefings last year. The feature, spotted in recent product builds, appears near release and would mark a shift toward AI assistants that anticipate user needs rather than waiting for requests.

Anthropic adds memory refinement and quality checks to managed agents platform.
Anthropic is rolling out “dreaming”—a system that reviews agent work sessions to identify patterns and improve performance over time—plus “outcomes,” which lets developers set success criteria that agents self-check against. Early customers like Harvey (legal work) and Wisedocs (document review) report significant gains: Harvey saw completion rates jump 6x, while Wisedocs accelerated reviews by 50%. The update signals Anthropic’s bet that agent productivity depends less on raw model intelligence and more on structured memory systems and feedback loops that keep agents aligned with business goals.

Anthropic and Wall Street giants form $1.5 billion AI consulting joint venture.
Anthropic partnered with Blackstone, Hellman & Friedman, and Goldman Sachs to embed its Claude AI and engineering teams inside mid-sized companies’ core operations, addressing a gap where companies lack in-house expertise for AI deployment. The move signals that AI labs now view private equity portfolios—representing trillions in assets across thousands of companies—as a critical new distribution channel, with Anthropic betting that enterprise AI adoption is bottlenecked by implementation capability rather than model access alone.

Claude gains ability to control your computer directly and autonomously.
Anthropic has released a research preview letting Claude operate computers like a human would—opening applications, browsing websites, and completing spreadsheet tasks without human intervention. This marks a shift from AI that answers questions to AI that can independently execute workplace tasks, currently available on macOS through Claude Cowork and Claude Code. The capability is significant because it moves AI assistants closer to handling routine office work end-to-end, though the preview status suggests the technology still needs refinement before broader release.

Google releases faster Gemma 4 models with triple speed boost.
Google released Multi-Token Prediction drafters for its Gemma 4 models, achieving up to 3x faster inference without sacrificing output quality or reasoning ability. The technique, called speculative decoding, pairs Gemma 4 with lightweight drafts that predict multiple tokens simultaneously, reducing the memory bottleneck that typically slows down AI models on consumer hardware. The free, open-source drafters are immediately available and work across devices from smartphones to workstations, directly addressing a core pain point for developers deploying AI applications.

Google prepares unified video generation tool ahead of I/O 2026.
Google is testing a new model called Omni that could combine video and image generation within Gemini, replacing its current split approach using separate Veo and Nano Banana systems. The move matters because it would simplify Google’s AI media tools and directly compete with ByteDance’s Seedance, which currently leads video-generation benchmarks. UI leaks suggest the tool may launch at Google I/O on May 19–20, 2026, though whether Omni represents a new model or integration wrapper remains unconfirmed.

Harvey releases first open-source benchmark for legal AI agents.
Harvey’s Legal Agent Benchmark (LAB) contains 1,200 real-world legal tasks across 24 practice areas, designed to measure whether AI agents can handle complex, multi-document legal work that mirrors how law firms actually operate. Unlike earlier legal AI tests that evaluated simple question-answering, LAB requires agents to navigate messy files, synthesize information across documents, and produce review-ready work product—with no partial credit, reflecting how high-stakes legal work is actually judged in practice.

Space-based solar power could transform AI’s energy demands.
Meta is investing in orbital solar technology through Overview Energy, which plans to launch thousands of satellites to beam power continuously to data centers—eliminating the need for batteries or conventional backup systems. This matters because AI data centers consume enormous amounts of electricity, and continuous space-based solar could reduce their reliance on grid power and fossil fuels. The approach remains experimental, but if viable, it would address one of AI’s thorniest infrastructure challenges: reliable, clean power at scale.

Meta bets billions on space solar and multi-day energy storage.
Meta has secured capacity reservations for two experimental energy technologies to power its AI data centers: satellites beaming solar energy from orbit to earth-based farms, and fuel-cell systems storing renewable power for over 100 hours instead of a few hours. Both partnerships target 2028 demonstrations with potential grid deployment by 2030, representing a shift from incremental renewable energy purchases toward backing early-stage infrastructure innovations that could fundamentally change how AI facilities secure reliable power at scale.

Figure AI’s humanoid robot masters stairs using only its camera.
Figure’s F.03 robot can now navigate stairs autonomously using only visual input from its onboard camera, without relying on additional sensors or pre-programmed instructions. The company trained the system end-to-end through simulation-based reinforcement learning, then deployed it to real robots—demonstrating that their humanoids can walk themselves from the factory floor to headquarters without human intervention. This marks a practical shift from controlled lab environments to functional autonomous navigation in real-world settings.

Pentagon signs deals with seven major AI companies for classified military networks.
The Department of Defense formally partnered with SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, and Amazon to deploy advanced AI systems on its classified networks, with Scale AI separately winning a $500 million contract. This marks a significant shift toward integrating frontier AI—the most capable, cutting-edge models—directly into sensitive military operations rather than keeping them at arm’s length, signaling confidence that these companies’ systems are ready for high-stakes national security use.

Meta acquires robotics startup to accelerate humanoid robot development.
Meta has acquired Assured Robot Intelligence (ARI), a startup specializing in AI that helps robots understand and adapt to human behavior, joining Meta’s AI research division. The acquisition signals Meta’s serious investment in humanoid robotics—a field many AI researchers believe is crucial to developing artificial general intelligence—though the company has yet to commit to a consumer product. Industry forecasts for the humanoid robotics market range from $38 billion to $5 trillion, reflecting both the technology’s potential and significant uncertainty around timelines.

ChatGPT integration brings AI formula writing and data analysis directly into Excel and Google Sheets.
OpenAI’s ChatGPT is now embedded as a spreadsheet add-on, letting users analyze messy data, generate formulas, and modify sheets without switching applications. This matters because it removes friction from routine data work—traditionally a context-switching hassle for office workers. The real distinction here is deep integration into existing tools rather than requiring a separate AI interface.

OpenAI secures $10 billion for new AI infrastructure venture.
OpenAI has launched a new company to build AI computing infrastructure, backed by investment firms TPG, Bain Capital, and SoftBank with $10 billion in committed capital. The move signals that major AI companies now need dedicated, massive investments in physical data centers and chips to stay competitive, rather than relying solely on cloud providers. The venture represents a shift toward vertical integration—controlling the entire supply chain from computing power to software—a strategy typically seen in capital-intensive industries.

OpenAI launches three new voice models for real-time conversations.
OpenAI released GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper in its API, enabling voice applications that reason, translate across 70+ languages, and transcribe in real time. GPT-Realtime-2 significantly improves over its predecessor with better instruction-following (70.8% vs. 36.7%), ability to handle tool-calling and interruptions, and GPT-5-level reasoning that keeps conversations flowing naturally. Early adopters like Zillow report 26-point improvements in call success rates, signaling that voice is becoming a viable interface for complex business tasks beyond simple voice commands.

Perplexity embeds licensed financial data directly into AI agents and workflows.
Perplexity is integrating professional-grade financial datasets—including real-time market data and research from providers like Morningstar and PitchBook—into its AI platform through both an API for developers and a dedicated finance product. This move distinguishes Perplexity by offering verifiable, licensed sources rather than relying solely on web search, addressing a key gap for finance teams that need current, cited information for investment decisions and analysis.

AI model performance hinges more on how you use it than the model itself.
How you instruct an AI system—through prompts, available tools, and supporting software—can dramatically shift its results on identical tasks, sometimes more than upgrading to a newer model. This matters because it suggests companies optimizing AI systems should focus heavily on engineering the “harness” around models rather than always chasing the latest, larger AI, potentially saving time and money while improving real-world performance.

Open-source robot model outperforms proprietary systems on real tasks
AI2 released MolmoAct 2, an open-source robot control model that beats closed commercial systems like Physical Intelligence’s π0.5 on real-world manipulation tasks, achieving 87% success on varied activities like moving objects and folding towels. The breakthrough matters because it makes capable robotics accessible to researchers—AI2 published complete code, 700 hours of training data, and a hardware reference design—rather than keeping recipes proprietary, potentially accelerating progress on the physical-world AI frontier where robots currently struggle with consistency.

Anthropic and OpenAI launch billion-dollar ventures for enterprise AI deployment
Both AI leaders are raising capital through joint ventures with major investment firms to embed AI agents directly into corporate workflows, with Anthropic’s $1.5 billion venture and OpenAI’s $10 billion venture operating on similar playbooks: securing preferred access to investors’ portfolio companies while deploying specialized engineering teams to customize AI tools for specific industries. This signals a strategic shift from selling generic AI products to businesses toward installing custom-built AI systems that integrate with existing corporate operations, mirroring Palantir’s successful model of embedded engineering teams.

Anthropic launches research institute to study AI’s real-world economic and security impacts
Anthropic has created The Anthropic Institute to investigate how AI affects jobs, markets, security, and innovation from within its own operations, then publish findings to inform public policy. The institute will release granular economic data, study dual-use risks from AI capabilities, and examine how AI changes work and institutions, positioning itself as an early warning system for significant disruption in an otherwise opaque frontier AI landscape.

Anthropic poised to launch new Claude model at May developer conference.
Anthropic has begun internal safety testing of a new AI model codenamed Jupiter V1, following the same pre-launch pattern used before releasing Claude 4 last year. The timing—weeks before the May 6th Code with Claude conference in San Francisco—suggests a new model announcement is imminent, likely to refresh Anthropic’s mid-tier offerings or introduce an entirely new generation. This matters because it signals continued competition in the race for more capable AI systems, with major labs announcing new models roughly annually.

Firefox security team fixes fifteen months of bugs in one month.
Mozilla’s Firefox development team used Claude Mythos Preview—an AI tool—to identify and fix more security vulnerabilities in April than they had in the previous 15 months. This represents a major acceleration in bug-fixing productivity that could have immediate implications for browser security across millions of users, demonstrating a concrete use case where AI tools measurably speed up critical defensive work.

AI agents need meaningful human oversight to stay competitive.
As AI systems like Claude become widely available, companies can’t compete by simply automating workflows—they need to deliberately preserve high-value decisions for humans. The real advantage lies in how firms design agent-human collaboration to keep people engaged in consequential choices, not in the AI tool itself.

Grok’s independence as a frontier AI model appears increasingly uncertain following recent deal activity.
Elon Musk’s xAI has been exploring strategic partnerships that suggest Grok may not compete at the highest tier of AI development on its own. The move signals potential constraints on xAI’s ability to maintain a cutting-edge model without external support or resources, undermining earlier positioning of Grok as a standalone frontier competitor to models like GPT-4 and Claude.

Apple settles AI feature lawsuit for $250 million without admission.
Apple agreed to pay up to $250 million to iPhone 15 and 16 buyers who sued over misleading claims about AI capabilities—specifically an enhanced Siri assistant that wasn’t available when promised. The settlement resolves a class action that accused Apple of false advertising to compete in the AI race, though Apple denied wrongdoing. The payout represents a rare instance of a major tech company facing financial consequences for overstating AI features rather than just adjusting marketing language.

Apple will let iPhone users choose which AI model powers their phone
Apple is planning to let iOS 27 users select from multiple AI models—including options from Google and Anthropic alongside the current ChatGPT integration—to power features like Siri and writing tools. This shift matters because it signals Apple’s strategy to position itself as a platform for AI rather than building proprietary AI services, addressing perceptions that the company has lagged competitors while avoiding massive infrastructure spending.

AI agents now rank on code restructuring capabilities.
Anthropic’s Claude Opus 4.7 leads a new leaderboard measuring how well AI can reorganize existing code while preserving functionality—a specific software engineering task distinct from writing code from scratch. This matters because refactoring is a frequent, detail-intensive real-world developer activity where mistakes can cause system failures, making it a meaningful benchmark for AI coding usefulness.

Coinbase cuts workforce by fourteen percent citing market conditions.
Coinbase CEO announced a 14% workforce reduction today, citing the need to right-size the company amid volatile crypto markets and shifting business priorities. The layoffs signal that even well-funded tech companies are adjusting headcount expectations as growth assumptions moderate across the sector.

AI regulation will likely reflect professional power, not just capability.
Doctors, lawyers, and other high-status professionals have political influence and donor networks that will shape AI policy regardless of technical feasibility. This suggests AI’s actual deployment in regulated professions depends less on what the technology can do and more on what incumbent professionals allow through regulation—a distinctly political, not purely technological, outcome.

Sierra hits $15 billion valuation in $950 million fundraising round.
The AI customer service startup has now secured over $1 billion in total funding, positioning itself to compete for enterprise adoption across industries. This valuation jump reflects investor confidence that AI-powered customer service represents a major business opportunity, though Sierra faces competition from larger tech companies moving into the same space.

AI chips worth billions are reaching China despite U.S. export controls.
The U.S. has attempted to restrict advanced chip sales to China, but new analysis suggests 290,000 to 1.6 million high-end processors are bypassing these controls—potentially representing one-fifth to three-fifths of China’s total AI computing capacity. This matters because it undermines U.S. technological dominance strategy and could accelerate China’s AI capabilities in ways policymakers didn’t anticipate, raising questions about enforcement effectiveness.

Google launches $3.5 million film competition for optimistic AI-powered futures
Google is partnering with XPRIZE to fund a global competition inviting filmmakers to create short films envisioning positive, technology-forward futures, with submissions open through August 2026. The competition aims to lower production barriers for emerging creators by offering the grand prize winner full production support to expand their three-minute film into a feature, using AI tools as part of the creative toolkit. This initiative reflects growing industry interest in shaping public narratives around technology through storytelling rather than technical development alone.

Google DeepMind partners with newly independent EVE Online studio.
CCP Games, rebranded as Fenris Creations after a $120 million management buyout from Pearl Abyss, is partnering with Google DeepMind to use its complex multiplayer game as a testing ground for AI systems that learn long-term planning and decision-making. DeepMind will run experiments on an offline version of the game to study how AI behaves in player-driven environments, while the partnership may also create new gameplay experiences. This partnership reflects how game worlds—which simulate complex human behavior at scale—have become essential infrastructure for advancing AI capabilities.

Google’s AI agent cuts DNA sequencing errors by thirty percent.
AlphaEvolve, a coding agent powered by Google’s Gemini model, improved DeepConsensus—software that corrects errors in DNA sequencing—reducing variant detection mistakes by 30%. The breakthrough enables scientists to analyze genetic data more accurately and cheaply, potentially uncovering disease-causing mutations previously missed, demonstrating how AI agents can solve real-world problems beyond software engineering.

Gemini’s lightweight model cuts costs for routine AI work.
Google released Flash Lite, a stripped-down AI model designed for high-volume, repetitive tasks like translation and data processing rather than complex reasoning. The move signals a shift toward cheaper, task-specific AI for businesses running thousands of daily operations, contrasting with the push toward more powerful general-purpose models that dominate headlines.

Google rebrands Fitbit app with AI health coaching features.
Google is merging Fitbit into a new Google Health app launching May 19, powered by an AI coach built on Gemini that offers personalized workout and health insights. The move consolidates Google’s health platforms and introduces a lightweight Fitbit Air tracker, while promising users that health data won’t feed into ad targeting—a notable privacy commitment as tech companies increasingly monetize health information.

Google releases Workspace command-line tool for AI agents.
Google launched gog 0.16, a command-line interface letting both humans and AI agents control Workspace apps like Gmail, Drive, Docs, and Sheets directly. The update adds safety features including sanitized data outputs and restricted command profiles, suggesting Google is preparing enterprise tools for autonomous AI use rather than human-only interfaces. This reflects a shift toward AI agents performing office work within existing corporate software ecosystems.

Google adds screen sharing and custom agents to its AI coding assistant.
Google is expanding Antigravity, its AI-powered IDE launched last year, with two capabilities currently being tested: screen sharing that lets developers show agents what’s happening outside the editor (useful for mobile debugging and live demos), and custom agent scripts that teams can build and invoke for specific tasks. These additions address long-standing developer complaints that Antigravity lacked visibility into external work and customization options compared to competitors like VS Code with Copilot, positioning it more competitively as a platform for coordinating multiple AI agents on coding projects.

Google expands Gemini API to search images and text together
Google has upgraded its Gemini API File Search tool to process images and text simultaneously, helping developers build systems that can find information across both photo and document archives using natural language descriptions. The update also adds custom metadata filtering to reduce irrelevant results and page-level citations so users can verify where answers came from—addressing a key need for enterprise applications requiring accuracy and transparency.

Maryland becomes first state to ban algorithmic grocery price hikes.
Maryland’s new law prohibits retailers from using AI systems to automatically raise prices on essential groceries, addressing consumer concerns about surge pricing in food shopping. The legislation is the first of its kind in the U.S. and signals growing regulatory scrutiny of algorithmic pricing in sectors considered vital to household budgets, though retailers argue such systems optimize inventory and reduce waste.

White House explores pre-release security checks for AI systems.
The Biden administration is considering requiring AI developers to submit models for government vetting before public release, aiming to catch safety risks early. This marks a shift toward proactive oversight rather than responding to problems after deployment, distinguishing it from current light-touch regulation. The proposal reflects growing concern that AI’s dual-use capabilities—beneficial applications alongside potential harms—warrant upstream intervention similar to pharmaceutical or nuclear oversight models.

Meta’s Hatch agent integrates AI automation with social media platforms
Meta is preparing Hatch, an autonomous AI agent launching via waitlist that can generate images, research topics, and handle shopping—with a distinctive twist of embedding these capabilities directly into Instagram and Facebook rather than a separate chat interface. The approach reflects Meta’s bet that users won’t switch platforms for AI tools, instead expecting agents to operate within their existing social feeds. Internal testing is targeted for late June, with a companion shopping agent planned for Instagram by year-end.

Codex surpasses Claude in downloads as enterprise developers embrace AI coding agents.
OpenAI’s Codex tool crossed ahead of Anthropic’s Claude Code on April 30, driven by rapid adoption of agentic capabilities—software that autonomously completes complex tasks like calendar management and code generation. The momentum appears tied to recent model upgrades (GPT-5.5 and Claude 4.7), with Codex doubling revenue in a week as enterprises prioritize practical AI tools over general-purpose assistants, signaling a market shift toward specialized, task-focused AI agents.

OpenAI’s former CTO testifies Altman lied about AI safety procedures.
Mira Murati testified under oath that CEO Sam Altman falsely claimed OpenAI’s legal department had approved skipping safety reviews for a new AI model, when in fact internal counsel disagreed with his statement. This marks another public accusation of dishonesty against Altman, joining similar claims from OpenAI cofounder Ilya Sutskever and board member Helen Toner, and underscores ongoing internal governance disputes as the company faces a high-stakes lawsuit from co-founder Elon Musk.

OpenAI open-sources MRC networking protocol for AI supercomputers.
OpenAI and five hardware partners released Multipath Reliable Connection (MRC), a new networking standard designed to keep hundreds of thousands of GPUs working reliably during AI model training. Unlike traditional networks that route data along single paths vulnerable to bottlenecks and failures, MRC spreads traffic across multiple simultaneous routes and recovers from link failures in microseconds—preventing the costly slowdowns that plague large-scale synchronized training. The protocol is already deployed across OpenAI’s largest supercomputers and is now available through the Open Compute Project for industry-wide adoption.

ChatGPT adds feature to alert trusted contacts during crises.
OpenAI launched Trusted Contact, a safety feature letting ChatGPT users designate someone to receive alerts if automated systems and human reviewers detect serious self-harm discussions. The feature addresses a real gap in AI safety by leveraging social connection—a clinically proven protective factor against suicide—while maintaining user privacy by sending only general notifications, not chat transcripts. The system undergoes human review within one hour and complements existing crisis hotline resources rather than replacing professional care.

ChatGPT gets a dedicated enterprise app for iPhone users.
OpenAI launched a separate iOS application designed specifically for business customers, complementing its consumer ChatGPT app. This move signals OpenAI’s strategy to build distinct products for different customer segments—enterprise clients get specialized features and controls while consumer users maintain their existing experience. The dual-app approach reflects how AI tools are increasingly becoming workplace infrastructure, not just consumer gadgets.

OpenAI launches faster ChatGPT model with major accuracy improvements
OpenAI released GPT-5.5 Instant as ChatGPT’s new default model for all users, featuring significant improvements in factuality—producing 52.5% fewer false claims on high-stakes topics like medicine and law compared to its predecessor. The free model now matches capabilities that paid competitors didn’t achieve until late 2025, with users reporting notably clearer, more concise answers alongside better image recognition and personalization features.

OpenAI releases framework for building multi-agent AI systems
OpenAI launched Agents SDK 2.0, an open-source toolkit that simplifies building workflows where multiple AI agents work together on complex tasks. The framework includes conversation memory, human oversight options, and safety guardrails—addressing practical challenges that developers face when coordinating multiple AI systems rather than just improving raw model capability.

OpenAI’s image generator learns to follow multi-step creative instructions precisely.
GPT-imagegen-2 can now execute complex, sequential requests—creating a 5×5 grid of progressively cuter dogs, then switching subjects to cats or squid, then generating book covers—in a single prompt. This marks a shift from earlier image models that struggled with nuanced instructions and object-specific modifications, suggesting AI is becoming more reliable for detailed creative work where precision matters as much as output quality.

I appreciate the task, but I’m unable to produce the requested summary. The material provided consists only of social media post titles, emoji tags, and broken or inaccessible links—with no actual content, context, or substantive information about what happened.
To write an accurate executive summary, I would need: – The actual news or announcement details – What these companies did or changed – Why it matters to business/society – Supporting evidence or data Could you provide the full articles, press releases, or detailed descriptions of these items?

AI search engine gains access to medical journals trusted by hospitals and doctors.
Perplexity’s new “Deep and Wide Research” feature now pulls answers directly from premium medical sources like the New England Journal of Medicine and British Medical Journal—the same publications hospitals and researchers use to make clinical decisions. This matters because it addresses a critical gap: AI health information is often unreliable, but sourcing answers from peer-reviewed journals adds credibility and traceability. The rollout includes nine additional medical databases coming soon, signaling a shift toward making AI-powered health research verifiable rather than just plausible-sounding.

China clears first commercial brain implant for paralysis patients
China has approved the NEO, a coin-sized brain chip developed by Shanghai-based Neuracle that reads electrical signals from the motor cortex to control a robotic glove, making it the world’s first invasive brain-computer interface cleared for commercial use. The device contains eight electrodes that detect when patients imagine moving their hands, then decode those signals to operate the glove in real time. This represents a regulatory milestone that could accelerate real-world deployment of neural implants beyond research settings, though questions remain about long-term safety, accessibility, and how other countries will respond.

AI models for drug discovery are getting smaller and more efficient than expected.
Stanford’s 2026 AI Index Report finds that compact protein-analysis models are now outperforming larger competitors on standard benchmarks, suggesting the field may be moving away from the “bigger is better” assumption that has dominated AI development. This matters because smaller models are cheaper to run and deploy in real-world medicine, potentially accelerating how quickly AI discoveries reach patients. The evidence: MSAPairformer, a relatively modest 111-million-parameter model, topped ProteinGym, the standard test for molecular biology AI tools.

xAI launches voice cloning tool for businesses and creators.
xAI now lets users create digital copies of their own voices from brief audio samples for use in customer service, content creation, and accessibility applications. The Custom Voices feature integrates with Grok’s text-to-speech and voice agent systems, allowing businesses to maintain consistent brand voices in automated customer support and enabling content creators to narrate at scale without repeated recordings. The capability marks a shift toward personalized AI voices replacing generic synthetic options across commercial applications.

xAI slashes Grok pricing by 40 to 60 percent
xAI released Grok 4.3, a notably cheaper version of its AI assistant that scores 53 on industry benchmarks—placing it competitively with Claude and other leading models. The dramatic price cut (40–60% lower costs) combined with improved agent capabilities makes it a significant play in the increasingly price-competitive AI market, where companies compete partly on cost-per-task rather than raw performance alone.