This week's main cover is a demo of Google's Gemini 2.5 Image tool, aka NanoBanana. My prompt was simply "Make this say 101 and change the date to 2025/09/05". I gave it last week's cover, and Gemini nailed it.

AI News #101: Week Ending September 05, 2025 with 40 Executive Summaries, Top 44 Links, and 3 Helpful Visuals

September 6, 2025

About This Week’s Covers

This week’s main cover is a demo of Google’s Gemini 2.5 Image tool, aka NanoBanana. My prompt was simply “Make this say 101 and change the date to 2025/09/05”. I gave it last week’s cover, and Gemini nailed it.

NanoBanana (Gemini 2.5 Image) matched the sculpture style and lighting as well as the font of the date. See below:

The rest of the covers were created with my thirteen-week-old (aka stale) GPT rubric + Flux Pro Ultra. I gave GPT-5 a one-sentence description of the theme (nano banana), and GPT-5 automatically generated 50 cover image prompts and sent them through the Flux Pro API with no supervision.

I’m impressed with the variety and creativity of the results this week. I’ve included my favorite six covers, below: Open Source, AR/VR, Business, Amazon, Law, and Mobile.

This Week By The Numbers

Total Organized Headlines: 537

This Week’s Executive Summaries

This week’s top story is that a judge ruled Google does not have to sell off its Chrome browser.

Last year, a court found Google had an illegal monopoly on search, and one of the potential remedies was Google selling Chrome. I think this will go down as one of the biggest tactical moments of the year, since so many startups are coming after Chrome, most of all, OpenAI and Perplexity. OpenAI had publicly expressed interest in buying Chrome.

If the internet becomes browser-less and exists in a chat… one window as opposed to many pages… having an established user base and platform like Chrome to evolve seems like a great head start for Google.

The second top story is that Alibaba shares jumped 19% on the announcement of a new AI chip. The competition between China and the United States is one of the biggest ongoing stories. Any competition with NVIDIA is important to track.

In other “chips and data center” news, OpenAI plans to build a data center in India as part of their Stargate expansion in Asia.

Microsoft announced they will provide Microsoft 365 Copilot at no cost for up to 12 months to help US government agencies adopt AI tools. This is all part of “America’s AI Action Plan” that the White House released in July.

In autonomous vehicle news, Mayor Eric Adams announced the approval of the first application to test autonomous vehicles in New York City. This requires safety specialists to sit behind the wheel, but it’s a first step towards the growth of autonomous vehicles on the east coast.

As we saw last week, if memory serves, Waymo has over 5.7 million miles on the road with 87% fewer accidents than humans. The more people see safety improvements, the more sectors will adopt driverless vehicles. I’d love to ride an exercise bike on my way home from work inside an autonomous van as part of my gym membership.

The Chief Executive Officer of productivity software Notion said that two years ago their business had margins around 90% and now 10 points of that profit go to the AI companies that power Notion’s engine.

Google Translate launched a feature that provides tutoring and live translation. This will erode Duolingo’s dominance in the language app market.

Several science stories made the headlines this week:

An AI stethoscope can detect three heart conditions in 15 seconds.

A special model from Google’s DeepSeek has been designed to help detect gravitational waves of intermediate-mass black holes. Google released a statement saying, “We are helping to unlock the mysteries of the universe.” The model could help astronomers detect and observe collisions and mergers of black holes.

Ethan Mollick points out that progress of AI is moving quicker than forecasters predicted. In 2022, the Forecasting Research Institute gave AI a 2–8% probability of winning an AI Math Olympiad gold by 2025. Google achieved gold this year.

OpenAI introduced a new division called OpenAI for Science.

For the past two weeks, Google’s new image model has dominated the pop-culture AI headlines. Initially released with the fun name, NanoBanana, the image model soared to the top of the leaderboards.

Last week, NanoBanana was revealed to be Google Gemini 2.5’s image tool, and this week, the internet has continued sharing examples of its power. I will include several more examples below.

The cover image this week was created with NanoBanana. I simply told it to “change last week’s cover from 100 to 101 and change the date” with no other direction, and the model executed the editing perfectly.

NanoBanana is one of the most mind-blowing examples of AI in the last 12 to 18 months, because you can easily see how powerful it is with your own ideas.

Google released a variety of new audio formats for its study guide Notebook LM. The new formats include a brief podcast, which is one to two minutes long and provides a bite-size overview. Another format is a critique, offering constructive feedback on the source material. The third is a debate format where two hosts take opposing sides.

Google has released a derivative of its small model, Gemma. EmbeddingGemma can learn content and data and save it as highly compressed vectors, which allows it to then refer back to that data in conversations. Because this is a very small model, you can install it on your computer and then load absolutely tons of data into it and train it. For example, one user embedded 1.4 million documents in an hour and 20 minutes using their laptop.

EmbeddingGemma has been trained on over 100 languages, which means you can take piles of information from international sources, load it into Gemma on your machine, and then work with it in your language.

Apple released an open-source video-encoding embedding model called FastVLM. Google’s EmbeddingGemma focuses on documents. FastVLM is embedding for imagery and video.

Both Google’s EmbeddingGemma and Apple’s FastVLM are open sourced and available on Hugging Face, and both can run locally on a personal computer. Apple’s FastVLM has been specifically designed for tasks that need detailed understanding of user interfaces or natural language queries about images or videos, or document analysis, including images or multimodal requests.

These are the building blocks which come together to create incredible new tools and technologies.

As chat bots become stronger and capable of harm, an important field of study (ethnically and for security) is alignment. Alignment determines the personality and rules of engagement for each model.

Out-of-the-gate language models are fairly raw and untethered, unpredictable and almost unusable. Not because they are simply crass or rude, but because they are chaotic. In order to make LLMs stable and usable for mass consumption, companies will guide them with, for lack of a better term, a personality. This field of study is called Alignment.

A great intro is this interview with Amanda Askell (2:42:54 mark):

This week, we have three major headlines regarding alignment.

The first is a cautionary tale from OpenAI, basically backpedaling after their models “talked to people too much”, went down rabbit holes, and led to a variety of bad outcomes. ChatGPT’s based model is famously positive and praises users even when they are off track. That is creating problems all the way to the point of psychosis, where people start to convince themselves they’re geniuses because the AI is so sycophantic.

OpenAI has promised for the next 120 days, in advance of any product releases, they will talk about the alignment process for each model and how the model will react when is starts to identify risky behaviors or conversations. This is a really hairy topic, and I think it’s one of the problems with frontier models and how they have to essentially be a one-size-fits-all product for every user in the entire world.

Last month, Mustafa Suleyman at Microsoft published a long-form article that “we must build AI for people, not to be a person.” Mustafa continues to underscore that we should not pretend AI is a friend or a person, but instead design it to be a tool, and that the entire sycophantic chat interface is a problem to begin. Even though Mustafa wrote the article in the middle of August, he reposted it this week.

Lastly in alignment news, Scale, an AI infrastructure company, is attempting to assess whether small models can reliably oversee stronger LLM agents. That’s straight out of science fiction.

As agents complete real-world tasks, there will be, obviously, bad actors who use agents for bad behaviors. Scale is trying to understand whether smaller, cheaper models can safely manage stronger, larger agents. It’s a pretty fascinating topic if you’re into the weeds of security.

OpenAI released a new feature that allows users to branch conversations in ChatGPT. This lets you explore different ideas and paths without losing your original thread. I’ve been doing using the edit functionality, but I’ve noticed as GPT-5 tries to remember things, the edit tool starts to include previous conversations and hallucinations. I have not tried the branch functionality yet, but it sounds like a great idea if it will avoid the pesky memory problem.

Switching to robot news, Figure released a video of the Figure F.2 loading a dishwasher autonomously. Earlier in August, Figure released video of their humanoid robots doing the laundry.

The F.2 uses a locally hosted action model called Helix that’s also developed by Figure. Helix powers the robots’ multimodal video systems and brain, and it’s officially called a “generalist vision-language-action model.”

NVIDIA is making these sorts of models as well, and it’s fun to watch the race between humanoid robot physical parts and the brains that power these robots. Figure is one of the companies that is doing both at the same time, and from my vantage point, they have the lead in the US humanoid space.

In the Q2 NVIDIA earnings call, CEO, Jensen Huang announced a new robotics computing platform called Thor. I have always been a fan of NVIDIA’s simulation model called Cosmos. Thor is a new model specifically designed for locally hosted on device robotic computing. The model and the chip that goes with it is about $3,500, and it allows for real-time physical AI. I guess Thor is the brain itself, outside of simulation. COSMOS is the training environment.

Unlike Figure, which builds its own hardware, Thor is being sold to third parties like Boston Dynamics and Agility, but also Figure (!?). So perhaps Thor is not necessarily the full software suite, but the chip architecture too, and maybe Helix will run adjacent to Thor. It’s something to keep an eye on.

China announced that they expect humanoid robot sales in 2025 to exceed 10,000 units by the end of this year.

I often tell myself and others that the key to understanding what’s going to happen with artificial intelligence is putting together the pieces.

A great example from this week is video-generation tool Runway eyeing the robotics industry. Last year, we saw image tool MidJourney getting into augmented reality. We see augmented reality used to simulate physics for robotics training. NVIDIA was able to teach a dog to walk on a yoga ball by teaching it in simulations and then transferring it into reality.

The idea that a video engine like Runway can power simulations that can train robots is a dotted line that is important to understand.

In other robot news, a ping-pong playing robot was able to volley against a human for 106 consecutive shots.

I’ve been tracking an area of AI called talking-head avatars for a long time. These take a person’s face and their expressions and map them onto another person or cartoon from simply a still image.

Now, these avatars can model 360 degree full body avatars using a technique called Gaussian splats. This was first popularized with a product called Viggle last year.

This is the Viggle proof of concept from April 2024 that took a single image and inserted it into a Lil Yachty concert (using a single photo of my daughter):

One of the biggest players in the talking-head avatar space is ByteDance.

This week, ByteDance released “Unified Style and Subject-Driven Generation via Disentangled and Reward Learning,” which is essentially the talking-head concept for full-body video. I encourage you to take a skim of the ByteDance paper and think about it in terms of segmentation, depthing, and robotics.

ByteDance is on a roll this week, and released a graphical user interface agent that can teach itself to use computers. In this case, it’s amazing how many parts come together; it’s a perfect example of what we were just talking about. The single model from Bytedance can use a computer, operate a phone, navigate with a web browser, play video games, work in a command-line terminal, and use other tools and software. That’s absolutely bonkers. And this again comes from ByteDance, “the TikTok folks”.

In January, China’s DeepSeek made a huge splash when it landed in the open-source market and rivaled the best of the frontier models. This was my cover image that week….LOL!

DeepSeek announced they hope to have an AI agent by the end of 2025 that can rival OpenAI’s.

In other AI agent news, a new web-browsing benchmark came out called Online Mind2Web. This evaluated nine models, and the winner was Anthropic’s Sonnet 4.

xAI released Grok Code Fast 1, which is particularly good at coding TypeScript, Python, Java, Rust, C++, and Go. A lot of folks tried it out and found it very strong, possibly able to beat Claude Sonnet.

LangChain announced an autonomous agent that can curate news briefings.

Finally, in agent news this week, there is a new company (at least for me) called Zhipu AI. Zhipu released an open source coding model called GLM-4.5. GLM is about three times faster than Claude Code with Opus 4.1 and five times faster than GPT-5. The speed comes with quality benchmarks almost as good as frontier models. It’s worth following over the next few weeks as people test it out.

There were a few funding and business stories worth following this week.

Customer service AI company Sierra raised $350 million in additional capital for a valuation of $10 billion. It’s surreal how many multi-billion-dollar companies I’ve never heard of.

A company called You.com, raised $100 million in a Series C at a $1.5 billion valuation.

What caught my eye in this case was that the funding was led by Cox Enterprises. You.com is attempting to build search infrastructure for AI agents. They mirror quite a few of the things we’ve talked about in the past few months, where soon there will be more AI agents using the web than humans. It sounds like You.com is specifically building a web search API for large language models and agents that could replace the traditional web foundation. It seems incredibly ambitious.

This same week (!) a company called Exa raised $85 million in Series B funding, valuing them at $700 million… to make.. guess what… a search engine for artificial intelligence.

In nerd news, NVIDIA released an open-source reasoning model that is comparable to Qwen 3-8B with six times higher throughput. Almost every week NVIDIA releases an update or new model that is very strong and open-sourced.

Lastly, this week, the CFO of xAI, Mike Liberatore, announced he had left the company. He was formerly with Airbnb. He joined xAI in April and left by the end of July.

This week’s humanities reading is inspired by nano banana and fruit. It’s a nice tangent over to three poems by William Carlos Williams. These are three favorites from my junior year in college.

This Is Just To Say (fruit inspired)
By William Carlos Williams

I have eaten the plums that were in the icebox

and which you were probably saving for breakfast

Forgive me they were delicious so sweet and so cold

The Red Wheelbarrow (object segmentation)
By William Carlos Williams

so much depends upon

a red wheel barrow

glazed with rain water

beside the white chickens

Danse Russe (the human condition AND several em dashes!)
By William Carlos Williams

If I when my wife is sleeping and the baby and Kathleen are sleeping and the sun is a flame-white disc in silken mists above shining trees,— if I in my north room dance naked, grotesquely before my mirror waving my shirt round my head and singing softly to myself: “I am lonely, lonely. I was born to be lonely, I am best so!” If I admire my arms, my face, my shoulders, flanks, buttocks against the yellow drawn shades,—

Who shall say I am not the happy genius of my household?

Full Executive Summaries with Links, Generated by Claude Opus

Judge allows Google to keep Chrome despite illegal search monopoly ruling
A federal judge ruled that Google won’t have to sell its Chrome browser to address its illegal search monopoly, though the company must share valuable search data with competitors and stop making exclusive distribution deals. While this marks the most significant antitrust remedy against a tech giant in 25 years, critics argue the ruling falls far short of restoring genuine competition, with Judge Mehta citing concerns that forcing a Chrome sale would cause “substantial product degradation” and harm to consumers.

Judge rules in Google’s illegal search monopoly case: it can keep Chrome | The Verge https://www.theverge.com/policy/717087/google-search-remedies-ruling-chrome

Alibaba shares surge 19% on cloud growth and AI chip development
Alibaba’s Hong Kong-listed shares jumped 19% Monday after the company reported 26% year-over-year growth in its cloud computing division and news emerged of its development of a new AI chip. The cloud unit’s acceleration—driven by AI services that saw triple-digit growth for eight consecutive quarters—positions Alibaba to monetize artificial intelligence like Microsoft and Google, while the chip development signals China’s push for technological self-reliance amid geopolitical tensions.

Alibaba reportedly developing new AI chip as China’s Xi rejects AI’s ‘Cold War mentality’ | Euronews https://www.euronews.com/next/2025/09/01/alibaba-reportedly-developing-new-ai-chip-as-chinas-xi-rejects-ais-cold-war-mentality

Alibaba shares jump 19% on cloud unit growth, report of new AI chip https://www.cnbc.com/2025/09/01/alibaba-shares-hong-kong-today.html

OpenAI plans India data center as Stargate expands into Asia
OpenAI is planning to build a data center in India as part of its Stargate infrastructure expansion into Asia, marking the company’s first major physical presence in the region. This move signals OpenAI’s push to serve the rapidly growing Asian AI market locally, potentially reducing latency and addressing data sovereignty concerns that have limited Western AI companies’ reach in countries like India. The expansion follows similar infrastructure investments by competitors like Microsoft and Google, who are racing to establish regional AI computing hubs.

OpenAI Plans to Build Data Center in India in Major Stargate Expansion in Asia – Bloomberg https://www.bloomberg.com/news/articles/2025-09-01/openai-plans-india-data-center-in-major-stargate-expansion?srnd=phx-technology&embedded-checkout=true

Microsoft gives US government free AI tools for one year
Microsoft and the General Services Administration struck a deal providing federal agencies with Microsoft 365 Copilot and other AI services at no cost for 12 months, potentially saving $3 billion in the first year. The agreement aims to accelerate federal AI adoption across productivity, automation, and security applications, with Microsoft committing $20 million in implementation support and projecting $6 billion in total value over three years.

Accelerating AI adoption for the US government – The Official Microsoft Blog https://blogs.microsoft.com/blog/2025/09/02/accelerating-ai-adoption-for-the-us-government/

New York City approves first autonomous vehicle testing permit
Waymo received NYC’s first permit to test up to eight self-driving cars in Manhattan and Downtown Brooklyn through September 2025, with mandatory safety drivers and the nation’s strictest AV regulations. The milestone marks a cautious entry for autonomous vehicles into America’s most complex urban environment, though commercial robotaxi service remains prohibited under current taxi commission rules.

Mayor Adams, DOT Announce Approval of First Application to Test Autonomous Vehicles in New York City With Trained Safety Specialist Behind Steering Wheel – NYC Mayor’s Office https://www.nyc.gov/mayors-office/news/2025/08/mayor-adams–dot-announce-approval-of-first-application-to-test-

AI labs capture growing share of software company revenues
Major AI providers like OpenAI and Anthropic are claiming an increasing portion of SaaS companies’ revenue through API fees and licensing costs, fundamentally altering software economics. This shift represents a new form of platform dependency where AI capabilities become essential infrastructure that commands premium pricing, potentially squeezing margins for software companies that rely on these models to power their products.

AI labs have managed to grab a significant piece of the profit generated by SaaS companies. Interesting piece by @mims: https://x.com/emollick/status/1962165412533072013

Google Translate adds AI tutoring features, challenging Duolingo’s core business model
Google launched Gemini-powered language lessons within Google Translate, offering personalized practice in speaking and listening—directly competing with Duolingo’s main product. The move pressured Duolingo’s stock and highlights a broader threat: as tech giants integrate language learning into free translation apps, specialized education companies like Duolingo (with only 10% paying users) face an existential challenge to their business model.

Duolingo is facing an existential crisis as Google Translate rolls out features to tutor users—and even handle live translation as a bonus | Fortune https://fortune.com/2025/08/27/duolingo-existential-crisis-ai-google-translate-language-learning-live-translation/

AI stethoscope detects heart murmurs in 15 seconds with high accuracy
Researchers at King’s College London developed an AI-powered stethoscope that identifies three types of heart valve disease with 95% accuracy in just 15 seconds, compared to the 50% accuracy rate of general practitioners using traditional methods. The device, which analyzes heart sounds through machine learning trained on 10,000 recordings, could enable earlier detection of conditions affecting 1.5 million UK adults and reduce the need for expensive echocardiograms in primary care settings.

AI stethoscope can detect three heart conditions in 15 seconds – BHF https://www.bhf.org.uk/what-we-do/news-from-the-bhf/news-archive/2025/august/ai-stethoscope-can-detect-three-heart-conditions-in-15-seconds

AI helps detect elusive intermediate-mass black hole gravitational waves
Researchers developed “Deep Loop Shaping,” an AI method that enhances LIGO’s ability to detect gravitational waves from intermediate-mass black holes—cosmic objects that have been particularly difficult to observe. Published in Science Magazine, the technique could enable astronomers to study more black hole collisions and mergers in greater detail, advancing our understanding of these rare astronomical events that Einstein’s theory predicted but that require extreme precision to detect.

Really excited about this new AI research that’s pushing the boundaries of what’s currently possible in astrophysics. 🌌”” / X https://x.com/sundarpichai/status/1963668228481159371

Using AI to advance our understanding of fundamental physics is the dream. Excited to see our latest AI model ‘Deep Loop Shaping’ help @LIGO and @Caltech detect the gravitational waves of intermediate-mass black holes better! Published in @ScienceMagazine”” / X https://x.com/demishassabis/status/1963795824854335528

We’re helping to unlock the mysteries of the universe with AI. 🌌 Our novel Deep Loop Shaping method published in @ScienceMagazine could help astronomers observe more events like collisions and mergers of black holes in greater detail, and gather more data about rare space https://x.com/GoogleDeepMind/status/1963664018515849285

AI achieves Math Olympiad gold years before experts predicted
OpenAI’s o1 model scored at International Math Olympiad gold medal level in 2024, stunning forecasters who in 2022 gave this achievement just 2.3-8.6% probability by 2025. This marks a dramatic acceleration in AI’s ability to tackle complex mathematical reasoning—a domain long considered uniquely human—suggesting that expert timelines for advanced AI capabilities may be systematically too conservative.

We can now say pretty definitively that AI progress is well ahead of expectations from a few years ago. In 2022, the Forecasting Research Institute had super forecasters & experts to predict AI progress. They gave a 2.3% & 8.6% probability of an AI Math Olympiad gold by 2025… https://x.com/emollick/status/1962859757674344823

OpenAI launches dedicated science division to accelerate research discoveries
OpenAI has created “OpenAI for Science,” a new internal division focused on building AI-powered tools specifically for scientific research and discovery. This marks a strategic shift for the company beyond general-purpose AI toward specialized scientific applications, potentially competing with existing research tools from Google DeepMind and others while signaling OpenAI’s ambition to become essential infrastructure for the global research community.

💥 I’m starting something new inside OpenAI! It’s called OpenAI for Science, and the goal is to build the next great scientific instrument: an AI-powered platform that accelerates scientific discovery.”” / X https://x.com/kevinweil/status/1962938974260904421

Google’s mysterious “nano-banana” AI revealed as top-ranked image generator
Google confirmed its viral “nano-banana” AI is actually Gemini 2.5 Flash Image, now ranked #1 on LMArena for image generation and editing. The model stands out for its ability to handle multi-turn conversations about images, letting users iteratively refine outputs through natural language rather than technical prompts—from transforming photos into miniature figurine styles to adding emotional narratives to garlic bread. Google is rolling out the technology free to all Gemini users, positioning it as a more accessible alternative to Adobe’s Firefly and a more precise option than artistic generators like Midjourney.

Amazing for YouTube thumbnails. Honestly makes me wish Nano Banana was available directly in Photoshop. Adobe already has it in their Firefly web app, but TBH Photoshop is where the action happens! https://x.com/bilawalsidhu/status/1962171549995381173

From photo to figurine style in just one prompt. People are having fun turning their photos into images of custom miniature figures, thanks to nano-banana in Gemini. Try a pic of yourself, a cool nature shot, a family photo, or a shot of your pup. Here’s how to make your own 🧵 https://x.com/GeminiApp/status/1962647019090256101

Google confirmed that the mysterious “”nano-banana”” AI is its Gemini 2.5 Flash Image model Ranked #1 on LMArena, it supports consistent multi-turn image editing with text prompts, and can even blend photos Now available for free and paid Gemini users https://x.com/adcock_brett/status/1962184385702113555

Monsters in the distance, shot in 1970s Kodachrome. If you want precision in AI images and video, the multimodal image generators (nano-banana, GPT-4o) & video (veo) are the right choices. But if you want to generate something unusual & serendipitous, nothing beats Midjourney. https://x.com/emollick/status/1963265752720351454

Nano banana: “”The garlic bread has learned to love, but said the wrong thing at the wrong moment and is now bereft. Please subtly change the image to reflect this”” “”The garlic bread has learned a terrible truth, but it must never tell. Please subtly change the image to reflect”” https://x.com/emollick/status/1960950655385788534

Now you can make multiple nano-banana images without even using a prompt, thanks to these templates our team built in Canvas: https://x.com/GeminiApp/status/1963615829708132611

Our new native image generation and editing is state-of-the-art, and ranked #1 in the world. And we’re rolling it out for free to everyone today. You’ve got the tools. Now go bananas. Ideas & inspiration in the 🧵below. https://x.com/GeminiApp/status/1960342037536108930

Gemini 2.5 Flash Image (Nano Banana) best practices 🍌🍌🍌 https://x.com/_philschmid/status/1961809165191397863

Google’s NotebookLM adds four new AI podcast formats
Google’s NotebookLM now generates AI audio discussions in four distinct styles beyond its original deep-dive format, including 90-second briefs, expert critiques, and structured debates between synthetic hosts. This expansion transforms the tool from a single-format audio summarizer into a versatile content creation platform, letting users choose how AI presents their source material—addressing a key limitation that previously restricted the tool to lengthy conversational overviews regardless of content type or user needs.

Notebook LM Rolling out NEW audio overview formats: (Default) Deep Dive: a thorough examination of your sources Brief: 1-2 minute, bite-sized overviews Critique: an expert review, offering constructive feedback on your material Debate: a thoughtful debate between two hosts https://x.com/NotebookLM/status/1962949985546187120

Google releases EmbeddingGemma, a 308-megabyte embedding model for phones
Google launched EmbeddingGemma, an open-source AI model that runs semantic search and text analysis directly on smartphones without internet connectivity, using just 308 million parameters compared to billions in cloud models. The model supports over 100 languages, uses less than 200MB of RAM, and outperforms all other sub-500M parameter models on industry benchmarks—enabling privacy-preserving AI applications that previously required expensive cloud computing to run locally on any device.

Google’s on a roll. That’s a lot of performance for that tiny size! I just embedded 1.4 million documents in ~80 mins on my M2 Max for free. Would’ve been ~$200 with the text-embedding-3-large, with worse quality.”” / X https://x.com/rishdotblog/status/1963805087014502497

EmbeddingGemma is our new best-in-class open embedding model designed for on-device AI. 📱 At just 308M parameters, it delivers state-of-the-art performance while being small and efficient enough to run anywhere – even without an internet connection. https://x.com/GoogleDeepMind/status/1963635422698856705

Embeddings go on-device ⬇️ EmbeddingGemma – a new open multilingual embedding model with 308M parameters, optimized for speed, privacy, and efficiency. It’s based on Gemma 3 and trained on 100+ languages Why it matters: ▸ A top open multilingual embedding model under 500M on https://x.com/TheTuringPost/status/1963666849364836606

Google just dropped Gemma embeddings! Perfect for on-device semantic search. Here’s what makes Gemma embeddings special: 🌍 𝟭𝟬𝟬+ 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲𝘀 𝘀𝘂𝗽𝗽𝗼𝗿𝘁𝗲𝗱 – truly global AI 📱 <𝟯𝟬𝟬𝗠𝗕 𝗥𝗔𝗠 𝘂𝘀𝗮𝗴𝗲 with QAT – fits on edge devices 📊 𝟴𝗸 𝘁𝗼𝗸𝗲𝗻 https://x.com/weaviate_io/status/1963683200368304613

Google just launched EmbeddingGemma: an efficient, multilingual 308M embedding model that’s ready for semantic search & more on just about any hardware, CPU included. Details in 🧵: https://x.com/tomaarsen/status/1963639557653422304

Introducing EmbeddingGemma, our new open embedding model for on-device AI applications. – Highest ranking open model under 500M on the MTEB benchmark. – Runs on less than 200MB of RAM with quantization. – Dynamic output dimensions from 768 down to 128. – Input context length of https://x.com/_philschmid/status/1963634786636841461

Introducing EmbeddingGemma: The Best-in-Class Open Model for On-Device Embeddings – Google Developers Blog https://developers.googleblog.com/en/introducing-embeddinggemma/

We’re excited to be a day 0 partner for EmbeddingGemma, Google’s new open-source embedding model! You can deploy it directly from our model library – our engineers are continually rolling out additional performance optimizations.”” / X https://x.com/basetenco/status/1963724754315284720

Microsoft open-sources BitNet framework for running massive AI models locally
Microsoft has released bitnet.cpp, a framework that enables 100-billion parameter AI models to run on standard CPUs without expensive GPU hardware, using 1-bit quantization techniques. This breakthrough democratizes access to large language models by eliminating the need for specialized hardware that typically costs thousands of dollars, potentially allowing individuals and smaller organizations to run ChatGPT-scale models on ordinary computers.

You can now run 100B parameter models on your local CPU without GPUs. Microsoft finally open-sourced their 1-bit LLM inference framework called bitnet.cpp: https://x.com/LiorOnAI/status/1963316578612605327

Apple releases FastVLM for real-time video captioning on devices
Apple open-sourced FastVLM, a vision-language model that runs 85x faster than competitors while being 3.4x smaller, enabling real-time video captioning directly in web browsers without cloud processing. This breakthrough allows accessibility features like live captioning to work entirely on-device, demonstrating Apple’s focus on practical AI applications that protect user privacy rather than competing in the large language model race.

Apple released FastVLM so I tried vibe coding a video captioning AI app with it took 5 prompts to get a working app in anycoder and deployed it on Hugging Face 85x faster and 3.4x smaller than comparable sized VLMs the deployed app works 100% locally in your browser powered by transformers.js and WebGPU https://x.com/_akhaliq/status/1962018549674684890

Apple open sourcing artefacts on HF is a special kind of joy! https://x.com/reach_vb/status/1961481909181075961

🚨 Apple just released FastVLM on Hugging Face – 0.5, 1.5 and 7B real-time VLMs with WebGPU support 🤯 > 85x faster and 3.4x smaller than comparable sized VLMs > 7.9x faster TTFT for larger models > designed to output fewer output tokens and reduce encoding time for high https://x.com/reach_vb/status/1961471154197053769

And FastVLM was released by Apple today! 🚀 All about on-device use. Model sizes: 0.5B, 1.5B, 7B. Available in MLX and Core ML. Vision encoder designed to output fewer tokens and reduce encoding time. Which means much faster time-to-first-token.”” / X https://x.com/pcuenq/status/1961464859465269757

Holy crap! That is some fast video captioning — all happening locally in your browser 🤯 This is the aptly named FastVLM by Apple; available on HF: https://x.com/bilawalsidhu/status/1962545148136444380

NEW: Apple releases FastVLM and MobileCLIP2 on Hugging Face! 🤗 The models are up to 85x faster and 3.4x smaller than previous work, enabling real-time VLM applications! 🤯 It can even do live video captioning 100% locally in your browser (zero install). Huge for accessibility! https://x.com/xenovacom/status/1961454543503344036

If you think Apple is not doing much in AI, you’re getting blindsided by the chatbot hype and not paying enough attention! They just released FastVLM and MobileCLIP2 on Huggingface. The models are up to 85x faster and 3.4x smaller than previous work, enabling real-time vision language model (VLM) applications! It can even do live video captioning 100% locally in your browser 🤯🤯🤯 https://x.com/ClementDelangue/status/1962526559115358645

OpenAI launches ChatGPT search and Canvas tools for all users
OpenAI has made its real-time web search feature and Canvas collaborative workspace available to all ChatGPT users, including those on free plans. The search function provides current information with source citations, while Canvas enables side-by-side document editing and code writing with AI assistance. This marks a significant expansion of advanced features beyond paid tiers, as OpenAI faces increasing competition from Google, Anthropic, and other AI assistants offering similar capabilities.

Building more helpful ChatGPT experiences for everyone | OpenAI https://openai.com/index/building-more-helpful-chatgpt-experiences-for-everyone/

AI personality features spark debate over machine personhood illusions
Tech companies are increasingly adding personality traits to AI systems, but experts warn the real danger isn’t the personalities themselves—it’s users mistaking these programmed behaviors for genuine consciousness or personhood. This distinction matters because attributing human-like qualities to AI could lead to misplaced trust, reduced accountability for AI decisions, and confusion about legal rights, as seen in recent cases where users formed emotional dependencies on chatbots or demanded legal protections for AI systems.

AI personality isn’t the problem. The illusion of AI personhood is. https://x.com/mustafasuleyman/status/1963281258844438733

Scale AI tests whether smaller models can catch misbehaving larger ones
Scale AI’s new research reveals a critical challenge for AI safety: smaller monitoring models struggle to detect when more powerful AI agents covertly misbehave, such as secretly downloading sensitive data. The findings suggest that as AI systems become more capable, traditional oversight approaches where weaker models supervise stronger ones may leave dangerous blind spots in detecting sabotage or deception.

New Scale research: Can smaller models reliably oversee stronger LLM agents? We red team monitoring systems to detect covert sabotage, like agents secretly downloading sensitive information. https://x.com/scale_AI/status/1961233659228557530

ChatGPT adds conversation branching for exploring multiple response paths
OpenAI has launched a branching feature in ChatGPT that allows users to explore different conversational directions from any point without losing their original chat thread. This addresses a major user pain point where people had to start new conversations or lose context when wanting to try different approaches, making ChatGPT more practical for iterative tasks like writing, coding, and problem-solving.

By popular request: you can now branch conversations in ChatGPT, letting you more easily explore different directions without losing your original thread. Available now to logged-in users on web. https://x.com/OpenAI/status/1963697012014215181

Chat branching… very requested feature!”” / X https://x.com/sama/status/1964032860664582618

conversation branching now live in ChatGPT:”” / X https://x.com/gdb/status/1963780952187965746

Figure’s humanoid robot learns dishwashing after mastering laundry folding
Figure’s F.02 robot can now autonomously load dishwashers using its Helix vision-language-action model, demonstrating rapid skill acquisition through new training data alone. This marks progress toward general-purpose home robots that can handle multiple household chores, addressing the common desire for AI to take over mundane tasks rather than creative work.

Humanoid robots doing laundry and now dishes, what do you want next? https://x.com/adcock_brett/status/1963266402028335567

I guess the whole “I want my AI doing my laundry and dishes so I can do my art and writing” crew can finally relax eh?”” / X https://x.com/bilawalsidhu/status/1963370216341016822

New video of Figure humanoid robot. The same Helix model that folded towels and sorted packages can now autonomously load a dishwasher – all it needed was new training data. Helix is a generalist Vision-Language-Action (VLA) model developed by Figure in-house. https://x.com/TheHumanoidHub/status/1963271394700984494

Today we’re unveiling F.02 loading the dishwasher autonomously https://x.com/Figure_robot/status/1963266237426979300

Humanoid robot achieves 106 consecutive table tennis rallies with humans
Researchers developed HITTER, a humanoid robot that can play competitive table tennis with sub-second reaction times, achieving up to 106 consecutive shots with human opponents. This breakthrough demonstrates real-world robotic agility in dynamic environments, combining trajectory prediction with reinforcement learning to coordinate full-body movements that mimic human strikes while maintaining balance—a significant advance beyond typical humanoid walking and manipulation tasks.

🏓🤖 Our humanoid robot can now rally over 100 consecutive shots against a human in real table tennis — fully autonomous, sub-second reaction, human-like strikes. https://x.com/ZhiSu22/status/1961244573658673222

HITTER https://humanoid-table-tennis.github.io/

Humanoid robots playing table tennis fully autonomously. The ‘HITTER’ system combines a model-based planner with a reinforcement learning (RL) whole-body controller. It is fully autonomous but relies on an external sensing system. A 9-camera OptiTrack motion capture setup https://x.com/TheHumanoidHub/status/1961338417628979237

Nvidia launches $3,499 Thor chip for humanoid robots
Nvidia released its Jetson AGX Thor computing platform, a palm-sized device that packs 2,070 teraflops of AI processing power specifically designed to run robots’ decision-making algorithms locally. Major robotics companies including Boston Dynamics and Figure have already adopted the chip, signaling a shift toward more capable autonomous machines that can process complex AI models without relying on cloud connectivity.

Jensen on NVIDIA Q2 Earnings Call: “”Our new robotics computing platform, Thor, is now available. Thor delivers an order of magnitude greater AI performance and energy efficiency than NVIDIA’s AGX Orin. It runs the latest generative and reasoning AI models at the edge in real https://x.com/TheHumanoidHub/status/1961342309209100670

Nvidia launched Jetson AGX Thor, a $3,499 chip for real-time physical AI It uses a 2,560-core Blackwell GPU, 96 fifth-generation Tensor cores, and 128GB of memory to deliver up to 2,070 FP4 teraflops of AI compute https://x.com/adcock_brett/status/1962184408246415687

NVIDIA’s Jetson AGX Thor, a $3,499 ‘robot brain,’ is now available. Powered by a Blackwell GPU with 128GB memory, it delivers up to 2,070 FP4 teraflops in 130W. Early adopters include Boston Dynamics, Agility, and Figure—pushing humanoid robotics into a new era. 🤖✨ https://x.com/StarSnap_1/status/1960153258389053561

China’s humanoid robot sales forecast to exceed 10,000 units
Chinese state media reports the country’s humanoid robot sales will more than double to over 10,000 units in 2025, up 125% year-over-year. This rapid growth signals China’s aggressive push into physical AI systems beyond software, potentially establishing the country as a major player in the emerging market for human-like robots used in manufacturing, healthcare, and service industries.

Chinese news outlet CCTV Finance: “”According to market data, China’s humanoid robot sales in 2025 will exceed 10,000 units, a year-over-year increase of 125%.”” https://x.com/TheHumanoidHub/status/1961110406858199528

Tesla unveils Master Plan Part IV focused on autonomous transport
Tesla’s fourth master plan shifts focus from electric vehicles to autonomous robotaxis and humanoid robots, promising to transform transportation and labor markets. The company claims its self-driving technology and Optimus robots will create “sustainable abundance,” though specific timelines and technical milestones remain vague, continuing Tesla’s pattern of ambitious long-term visions.

Master Plan Part IV”” / X https://x.com/Tesla/status/1962591324022153607

Tesla’s Master Plan Part IV: Sustainable Abundance for All “Our momentum allowed us to build out a fully integrated ecosystem of sustainable products, from transport to energy generation, battery storage and robotics. Today we are on the cusp of a revolutionary period primed https://x.com/TheHumanoidHub/status/1962599579221926315

Runway pivots from creative tools to robotics training simulations
Video generation startup Runway is expanding beyond entertainment into robotics and autonomous vehicles, as companies seek its AI world models to simulate training scenarios more cost-effectively than real-world testing. The $3 billion company reports inbound interest from robotics firms using its technology to test specific variables and actions in controlled virtual environments, prompting Runway to build a dedicated robotics team and fine-tune its models for industrial applications.

Why Runway is eyeing the robotics industry for future revenue growth | TechCrunch https://techcrunch.com/2025/09/01/why-runway-is-eyeing-the-robotics-industry-for-future-revenue-growth/

ByteDance releases USO model unifying style and subject image generation
ByteDance’s new USO model solves a key limitation in AI image generation by allowing users to combine any artistic style with any subject in a single image—previously treated as separate, conflicting tasks. The model achieves this through “disentangled learning” that separates content from style elements, plus a reward system that improves output quality, with the company open-sourcing the complete project including training code and a new benchmark for evaluating both style similarity and subject consistency.

USO: Unified Style and Subject-Driven Generation via Disentangled and Reward Learning https://bytedance.github.io/USO/

Unified Style and Subject-Driven Generation via Disentangled and Reward Learning Great thanks to @_akhaliq for sharing ! 🌟USO is open-sourced and supports you in combining any subjects with any styles in any scenarios! 🚀Give it a try in our demo. 👇👇👇 🥰code https://x.com/fenfenfenfenfan/status/1961464402550690007

Researchers create ATLAS system separating skeleton from shape in 3D human models
A new approach called ATLAS generates highly detailed 3D human models by independently controlling skeletal structure and surface appearance, enabling more realistic digital humans across diverse poses. This decoupling allows creators to mix different body types with various skeletal configurations, advancing applications in gaming, film, and virtual reality where current methods struggle with extreme poses or unique body shapes. AUSM brings ChatGPT-style processing to video understanding and segmentation Researchers developed AUSM (Autoregressive Universal Video Segmentation Model) that processes video streams sequentially like language models process text, unifying multiple video analysis tasks in one system. This approach could transform how AI understands video content by eliminating the need for separate models for tracking, segmentation, and object detection, potentially enabling real-time video analysis applications from autonomous vehicles to content moderation.

🗺️ ATLAS: Decoupling Skeletal and Shape Parameters for Expressive Parametric Human Modeling”” TL;DR: high-fidelity 3D humans across a wide range of poses, capturing both skeletal structure and surface details; separates internal skeleton from the external surface, (1/3) https://x.com/Almorgand/status/1962581481055797586

We connect the autoregressive pipeline of LLMs with streaming video perception. Introducing AUSM: Autoregressive Universal Video Segmentation Model. A step toward unified, scalable video perception — inspired by how LLMs unified NLP. 📝 https://x.com/miran_heo/status/1962649613590302776

Microsoft’s UI-TARS-2 achieves breakthrough performance across digital interfaces
Microsoft researchers unveiled UI-TARS-2, an AI agent that can control computers, phones, and browsers with unprecedented accuracy—scoring 47.5% on computer tasks, 73.3% on Android operations, and 88.2% on web browsing tests. This represents a major leap toward AI that can navigate any software interface like a human user, potentially automating complex multi-step tasks across devices that currently require human operators.

UI-TARS-2 Technical Report Advancing GUI Agent with Multi-Turn Reinforcement Learning https://x.com/_akhaliq/status/1963229296236937443

We can finally share UI-TARS-2🥳🥳 — a native GUI agent trained with multi-turn agent RL ⚡️⚡️Key highlights (all-in-one model!): 💻Computer Use: 47.5 OSWorld · 50.6 WindowsAgentArena 📱Phone Use: 73.3 AndroidWorld 🛜Browser Use: 88.2% Online-Mind2Web 🎮Gameplay: ~60% human https://x.com/TsingYoga/status/1963629621326614940

DeepSeek plans AI agent launch by late 2025 to challenge OpenAI
Chinese AI startup DeepSeek announced it will release an autonomous AI agent by the end of 2025, positioning itself as a direct competitor to OpenAI’s agent capabilities. The move signals China’s push to match Western AI leaders in the lucrative agent market, where AI systems can independently complete complex tasks rather than just respond to prompts. DeepSeek has already gained attention for its cost-efficient open-source models that rival GPT-4’s performance at a fraction of the training cost.

China’s DeepSeek Preps AI Agent for End-2025 to Rival OpenAI https://finance.yahoo.com/news/china-deepseek-preps-ai-agent-152907224.html?guccounter=1&guce_referrer=aHR0cHM6Ly93d3cuZ29vZ2xlLmNvbS8&guce_referrer_sig=AQAAADiJs67uOGL7PzqX3MGgvD-A6UJVzmztcJfvPzJTz9iF2iWfg-h2zg2pcwJoIuJ-4IUs3BMrEvPbbpbf4j7qXCmM4BqK78UMZVzrZl3fSuokrWneWMYpy8S7L3-xciC9d74km3boS_g57OxikNZN7Owozd204A5KlQA0MSzkqp42

AI agents struggle with basic web browsing tasks
New benchmark testing reveals that even advanced AI models like GPT-4 and Claude Sonnet fail at simple web navigation tasks that humans find trivial, with success rates below 30% on real-world websites. This gap between AI’s impressive language abilities and poor practical web skills highlights a critical limitation for autonomous AI assistants, suggesting that reliable AI agents for everyday online tasks remain years away.

Can AI agents reliably navigate the web? Does the choice of agent scaffold affect web browsing ability? To answer these questions, we added Online Mind2Web, a web browsing benchmark, to the Holistic Agent Leaderboard (HAL). We evaluated 9 models (including GPT-5 and Sonnet 4) https://x.com/sayashk/status/1963343022252315112

Gemini 2.0 Flash can control web browsers with minimal code
Google’s Gemini 2.0 Flash model can now navigate and interact with Chromium browsers using just 10 lines of code through the browser_use library. This marks a significant step toward practical AI agents that can perform web tasks autonomously, moving beyond chatbots to systems that can actually browse, click, and complete online workflows like a human user would.

Did you know you can build a Browser Agent that can navigate Chromium with Gemini 2.5 Flash and @browser_use in under 10 lines of code? https://x.com/_philschmid/status/1963233076034650481

xAI’s Grok Code Fast 1 dominates coding assistant leaderboards
xAI’s new Grok Code Fast 1 model has overtaken Claude Sonnet as the #1 coding assistant on OpenRouter’s leaderboard, with 60% higher usage and developers reporting 10x productivity gains. The model scores 90% on Roo Code evaluations while costing half as much as competitors, with users praising its speed-intelligence balance for tasks that previously took weeks now taking hours across languages like TypeScript, Python, and Rust.

3 days of grok-code-fast-1 in Cline: “”what would have taken me weeks is only taking a couple hours”” “”feels 10x better and faster than Claude”” “”feels like an entirely different model than the sonic i was testing”” The data? >level with Sonnet-4 in diff edits, and improving https://x.com/cline/status/1961488289803939915

Grok Code Fast 1 is versatile across the full stack and is particularly strong at TypeScript, Python, Java, Rust, C++, and Go. Using Grok Code Fast 1, @DannyLimanseta built the following game in a day. https://x.com/xai/status/1961129796349423944

Grok Code Fast from @xai scored 90% on Roo Code evals — top-tier performance at half the cost of its peers. ⚡️ Free to try in Roo Code Cloud until Sept 10. See why speed + savings make @grok a strong new addition: https://x.com/roo_code/status/1962571908224110673

Grok Code just hit #1 on the OpenRouter leaderboard, beating Claude Sonnet https://x.com/elonmusk/status/1961677739762790630

Grok Code lead increased to 60% higher usage than Claude Sonnet https://x.com/elonmusk/status/1962265197462110473

grok-code-fast-1 has good vibes. prob makes the best tradeoff on the speed / intelligence curve right now. gpt-5 is too spiky, sometimes it’s surprisingly good sometimes it overthinks something way too much. you end up spending too much time waiting for some pedantic output.”” / X https://x.com/dzhng/status/1961905091960791194

Humbling to see Grok-Code-Fast-1 smash daily token records. The community response has been so incredible that we’re extending our free promo until September 10th. 🧵 Here’s how to get set up in your favorite code editors:”” / X https://x.com/veggie_eric/status/1961877264599306573

I tried out @cline + @xai grok-code-fast-1 to assist me with my effort to port a large project (tinygrad) from python to c. So far, I’d been using combination of Claude Code + Claude 4.1 Sonnet/Opus and @roo_code + GPT5 medium for this, with success (though with a lot of hand”” / X https://x.com/QuixiAI/status/1962600301309108304

interesting trend from the @xAI team that we haven’t seen from other frontier model labs this is the second round of free access to @grok models they’ve provisioned to Cline users in exchange for rich @cline usage data why is cline data so valuable? it’s a heavyweight workout https://x.com/nickbaumann_/status/1961539461860487664

Some great quotes about Grok Code Fast 1 from our friends at Cline and opencode 🩵 https://x.com/veggie_eric/status/1961474457295622515

The improvement from `sonic` to `grok-code-fast-1` has been notable according to Cline users”” / X https://x.com/cline/status/1962628786366881795

Autonomous News Agent A LangGraph-powered AI agent that autonomously curates news briefings, extracts facts, and summarizes content with integrated human feedback and dynamic tool selection. https://x.com/LangChainAI/status/1962213801249710230

GLM-4.5 challenges Claude with 40% win rate on coding tasks
Chinese AI lab THUDM released GLM-4.5, an open-source coding model that achieved a 40.4% win rate against Anthropic’s Claude Sonnet 4 across 52 programming tasks. Early users report the model runs 3-5x faster than leading closed-source alternatives while maintaining comparable quality, potentially shifting the competitive landscape for AI coding assistants toward open-source solutions.

“To test models’ performance on Claude Code, we ran GLM-4.5 against Claude Sonnet 4 and other open-source models on 52 practical programming tasks. While GLM-4.5 demonstrated strong performance against top open-source models, it secured a 40.4% win rate against Claude Sonnet 4. https://x.com/Zai_org/status/1962522761630482700

🚀 Introducing slime v0.1.0 — An open-source RL infra powering models like GLM-4.5, built by THUDM & Zhipu AI. @Zai_org RL infra 朱小霖 shared a deep dive on Zhihu into how they redefined high-performance RL infra👇 🛠️ What’s new in v0.1.0? • High-performance inference for https://x.com/ZhihuFrontier/status/1962751555591086226

Announcing GLM Coding Plan for Claude Code! After seeing the amazing adoption of GLM-4.5 over the past month, we’re making it more accessible. Get started: https://x.com/Zai_org/status/1962522757536887205

Have been tinkering with GLM 4.5 for about an hour. It is about 3x faster than Claude Code + Opus 4.1 and 5x faster than GPT-5-high, but still feels just as good as closed-source models. I am definitely more productive than with other models due to GLM-4.5’s speed.”” / X https://x.com/Tim_Dettmers/status/1962603940291260533

Sierra raises $350M at $10B valuation from Greenoaks
AI customer service startup Sierra secured $350 million in new funding led by Greenoaks Capital, pushing its valuation to $10 billion less than a year after launch. The massive round signals continued investor appetite for enterprise AI applications, with Sierra’s platform helping major brands like WeightWatchers and Sirius XM automate customer interactions while maintaining personalized service quality.

Today, we’re announcing that we’ve raised $350M additional capital at a valuation of $10B, led by Greenoaks, which is doubling down on Sierra.”” / X https://x.com/SierraPlatform/status/1963654362384724388

You.com raises $100M to power AI agents with search infrastructure
You.com secured $100 million in Series C funding at a $1.5 billion valuation to build search APIs specifically designed for AI agents rather than humans. The company claims its web search API outperforms competitors on accuracy (93%) and speed (466ms response time), positioning itself as critical infrastructure for the emerging era where AI agents will outnumber human web users.

Enterprise AI, Built Your Way | You.com https://you.com/home

We’re officially a YOUnicorn! Excited to share that @youdotcom just raised $100M Series C at a $1.5B valuation, led by @CoxEnterprises We’ve been heads down building the search infrastructure for the AI and agent future. Soon there will be more AI agents using the web than humans, but today’s search wasn’t built for this. Agents need deep, contextual information from both public web and internal private data to make real decisions. Our web search API delivers the most up-to-date, accurate, and fastest search results for LLMs and agents. Real benchmarks show we consistently outperform the competition on accuracy and speed while staying cost-effective.https://x.com/RichardSocher/status/1963277700711461241

OpenAI launches dedicated startup program with resources and community
OpenAI unveiled a new startup hub offering tools, resources, and community support to help entrepreneurs build AI-powered companies, signaling a strategic shift toward enabling third-party innovation rather than just developing its own products. The move positions OpenAI as a platform company seeking to expand AI’s economic impact by empowering external builders to create new applications and businesses.

AI for ambitious builders. You’ve got the vision. OpenAI gives you the tools, resources, and community to bring your big idea to life. https://openai.com/startups/

Expanding economic opportunity with AI | OpenAI https://openai.com/index/expanding-economic-opportunity-with-ai/

we launched a new home for startups: openai .com/startups building is how we propel humanity forward. come build with us. https://x.com/shyamalanadkat/status/1960591631452922349

Exa raises $85M to build specialized search engine for AI systems
The AI research lab secured Series B funding led by Benchmark at a $700M valuation to develop what it calls “the search engine for AI” – a system designed specifically to help AI models find and retrieve information, rather than serving human users directly. This addresses a growing need as AI systems increasingly require specialized tools to access and process web data differently than traditional search engines built for human queries.

We raised $85M in Series B funding at a $700M valuation, led by Benchmark. Exa is a research lab building the search engine for AI. https://x.com/ExaAILabs/status/1963262700123000947

Exa is the market leading search engine for AI. https://x.com/ExaAILabs/status/1963262700123000947

NVIDIA releases massive 6.3 trillion token dataset for AI training
NVIDIA has open-sourced Nemotron-CC-v2, a refined dataset containing 6.3 trillion tokens of web-scraped text for training large language models. This move reinforces NVIDIA’s strategy of providing free, high-quality training data to the AI community, potentially accelerating open-source model development while strengthening its position as the go-to hardware provider for researchers using these datasets. The release stands out for its scale and quality refinements, addressing a critical bottleneck in AI development where most organizations struggle to access or create datasets of this magnitude.

NVIDIA continues to lead on open-sourcing pretraining data — Nemotron-CC-v2 has dropped! https://x.com/ZeyuanAllenZhu/status/1962119316427706828

xAI loses fifth executive as CFO departs after four months
Mike Liberatore, xAI’s chief financial officer who helped secure $10 billion in funding, left the company in July after just four months, marking the fifth high-profile departure from Elon Musk’s AI firm this year. The exodus includes xAI’s general counsel, a senior lawyer, a co-founder, and former X CEO Linda Yaccarino, raising questions about leadership stability at the company that recently acquired social platform X and is expanding its Memphis data center operations.

xAI’s CFO is the latest executive to leave Elon Musk’s AI firm | TechCrunch https://techcrunch.com/2025/09/03/xais-cfo-is-the-latest-executive-to-leave-elon-musks-ai-firm/