Daily Posts on AI,Agent,SpaceX,OpenAI,Anthropic and more
SF
Joined June 2026
- Tweets2.5K
- Following39
- Followers106
- Likes1.5K
Pinned Tweet
Tesla is hiring a Machine Learning Engineer for Realtime Games on Digital Optimus in Palo Alto. The job is agents that play AAA games, browse the web and run pro apps at superhuman level.
Digital Optimus is the Tesla and xAI project Musk announced in March. It already gets about halfway through the Diablo campaign just by looking at the screen.
It is good at Counter Strike too, and League is in training.
Games are the cheapest stress test for computer use. Fast screens, messy inputs, instant feedback on every mistake.
An agent that survives a Diablo boss fight can survive a spreadsheet
Tesla is hiring a Machine Learning Engineer for Realtime Games, Digital Optimus, in Palo Alto. The job: build agents that play AAA games, browse the web and use pro apps at superhuman levels. Why games? Digital Optimus, the Tesla and xAI project Musk announced in March, reads screen video and acts with a keyboard and mouse. AAA games are a safe, endless, real-time proving ground for exactly that, the same approach DeepMind's SIMA uses. Tesla's posting says it will benchmark how skills transfer across games, general computer use, FSD and Optimus. tesla.com/careers/search/job…
Dario Amodei walked out of OpenAI in 2021 with his sister and a pitch that sounded like a handicap. Build AI safely and still win.
Five years later Anthropic hit a $965B valuation in May and an $81B revenue run rate by October.
Google and Nvidia each wrote $10B checks. Now the IPO is set at up to $1.4 trillion under the ticker ANT.
The pitch to Wall Street goes past chatbots. Amodei is telling investors cancer and Alzheimer's will be relics of the past in less than a decade.
The safety lab everyone called too careful to compete is about to list as one of the biggest companies on the market
> be Dario Amodei
> leave OpenAI in 2021 with your sister
> start Anthropic to build AI "safely"
> everyone says safety labs can't win
> build Claude
> $965B valuation in May
> $81B run-rate revenue by October
> Google and Nvidia write you $10B checks each
> go to Wall Street
> tell them cancer and Alzheimer's will be "relics of the past in less than a decade"
> set the IPO at up to $1.4 trillion
> ticker: ANT
1
5
CoreWeave reported earnings and the whole AI hardware stack woke up with it.
Sandisk closed up 13.7% on the day. Bloom Energy added 13.0% as AI infrastructure names rebounded on the back of those CoreWeave numbers.
Micron rose 11.9%, and the read there is simple. Strong results from a GPU cloud tell the market more memory is getting bought.
Three of the top movers in the US market today were storage, power and memory. None of them build models.
The money in this cycle keeps flowing to whoever sells the shovels, and one earnings report from a single neocloud was enough to move all of them
8
80k out of 2.68M. That is how many of Anthropic's chips come from Nvidia, about 3% of everything it runs or is bringing online.
The rest is Amazon and Google silicon. Roughly 1M Trainium2, an estimated 600k Trainium3 and an estimated 1M TPU v7.
Compare that with xAI and Microsoft at 100% Nvidia and OpenAI at 88%.
AMD is next in line with up to 2 GW of MI450 from the first half of 2027, which makes four chip vendors for one lab.
Everyone else is paying the Nvidia tax. Anthropic quietly built its way around it
Anthropic is the least Nvidia-dependent lab in AI.
Of the 2.68M chips it runs or is bringing online, 80k are Nvidia. 3%. The rest is Trainium and TPU.
xAI and Microsoft: 100% Nvidia. OpenAI: 88%.
Claude on AMD in 3 days just makes it four vendors.
The breakdown, live + tracked in progress as of Nov 1:
- Trainium2: 1.00M (Anthropic: "over one million")
- Trainium3: 600k (est.)
- TPU v7: 1.00M (est., Broadcom direct + Google Cloud)
- GB200: 80k (30k on Colossus 1, 50k Azure est.)
- Next: up to 2 GW of AMD MI450 from 1H 2027 (AMD).
flow.kydo.sh
12
Anthropic shipped a full workplace suite on Nov 10. Claude Sheets, Claude Whiteboards, Claude Meetings and Claude Data, four products in one drop.
Meetings is the one that stands out. It joins a Google Meet call as a visible participant, takes notes, drafts the action items and starts doing the work while the call is still running.
Sheets builds multi tab workbooks on its own. Whiteboards and Data round out the stack.
That is four direct shots at the daily tools people already pay Google and Microsoft for.
The model race was about benchmarks. This one is about owning the workday
13
OpenAI just put up its first $10 billion quarter and still lost more than it brought in.
Per the WSJ, third quarter revenue rose about 50% from the second quarter and landed near three times the level from a year earlier.
The adjusted operating margin got better. Profit is still nowhere close.
Growth is not the hard part for OpenAI anymore. Paying for the compute behind that growth is the whole game now
OpenAI just booked its first $10 billion quarter, and it still lost more than it made.
According to Wall Street Journal, third quarter revenue rose about 50% from the prior quarter to $10 billion, nearly three times the level from a year earlier. The adjusted operating loss widened to $10.9 billion from $10.1 billion, even as the adjusted operating margin improved to minus 109% from minus 150%.
CFO Sarah Friar said that the company now has an annualized revenue run rate of about $50 billion, that enterprise revenue doubled, and that paying subscriptions have topped 100 million users. OpenAI has 1.2 billion weekly active users, and revenue is now split evenly between enterprise and consumer.
Rival Anthropic reported about $17 billion of third quarter revenue and a $567 million operating loss. Those loss figures are not perfectly comparable, because Anthropic's number includes stock based compensation and OpenAI's adjusted loss does not.
I think this is the contrast investors will be staring at as Anthropic heads into its IPO but one company is already much closer to making the math work.
ALT Wall Street Journal headline: OpenAI Continues to Suffer Heavy Losses as Revenue Rises to $10 Billion
11
IREN says its Horizon 1 build came in 22% faster than the industry benchmark.
Horizon 1 is the first of four 50MW AI Cloud deployments for Microsoft. Four of them puts 200MW of IT capacity on one customer.
Starting civils to a 19k GB300 cluster tested under NVIDIA Exemplar took 14 months. The industry benchmark is 18 months for the building alone, with no GPUs inside.
Fourteen months from dirt work to live racks is the real story here. Power and construction are the bottleneck in AI now, not just chips.
The companies that control the land, the power and the build crew are the ones landing the Microsoft deals
1
8
AlphaGenome 2 was trained on more than 107 million individual cell profiles across 68 tissues.
The post reports causal variant identification at 30% versus 16%, and directional gene effect identification at 43% versus 23%.
It also says the model reproduced key aspects of an ETS2 finding in gut macrophages in about 30 minutes, compared with around 18 months of experimental research.
The takeaway is a 30 minute computational result beside an 18 month experimental timeline
AI is getting closer to understanding how human biology actually works.
Google DeepMind just introduced AlphaGenome 2, an advanced AI model trained on more than 107 million individual cell profiles across 68 tissues.
This is a big step towards understanding why diseases develop at the cellular level.
It can predict how DNA mutations influence gene activity across hundreds of different human cell types.
New model can even identified nearly 2× more causal genetic variants than its previous benchmark:
> 30% vs 16% in causal variant identification.
> 43% vs 23% in directional gene-effect identification.
It even computationally reproduced key aspects of an inflammatory bowel disease discovery involving the ETS2 gene and gut macrophages in approximately 30 minutes, compared with around 18 months of experimental research.
We're still in the early stages of this transformation.
The coming years could be extraordinary for biotechnology, longevity and human health.
And we're only getting started with AI-driven science.
4
Gemini 4 Carbon just landed at #2 and #3 on Paweł Huryn's bug hunt benchmark. Two real repos, 105 replanted bugs, the job is to find and fix as many as possible.
Fable 5.2 on max still leads with 58. It cost $139.65 to get there.
Gemini 4 Carbon on high scored 57.3 for $31.32. On medium it scored 56 for $24.95.
Sonnet 5.5 on max landed at 51.3 for $154.03. GPT 6 Astra on max scored 45 for $33.03.
Almost the same score as the leader for under a quarter of the bill. Google stopped being the punchline of the model race this week
1
10
Sam Altman sat in front of Samsung executives and admitted there has been some double counting of AI demand in the neocloud sector.
Asked about AI, memory chips and data centers over the next five to ten years, his answer was ups and downs in the short term.
The bigger line came next. OpenAI and Anthropic are the only two companies really driving substantial growth in demand, and he warned of a supply shock and a period of turbulence.
That is the head of one of the biggest compute buyers telling the biggest memory maker that the order book has some air in it
During a Q&A with Samsung executives, Sam Altman addressed concerns about a bubble in the neocloud sector, saying, “There has been some double counting of demand.”
Asked whether demand for AI, memory chips, and data centers could level off or grow more slowly over the next five to ten years, he replied, “There are ups and downs in the short term.”
“OpenAI and Anthropic are the only two companies really driving substantial growth in demand, so we could see a supply shock and a period of turbulence,” Altman said.
6
Sam Altman sat in front of Samsung executives and admitted there has been some double counting of AI demand in the neocloud sector.
Asked about AI, memory chips and data centers over the next five to ten years, his answer was ups and downs in the short term.
The bigger line came next. OpenAI and Anthropic are the only two companies really driving substantial growth in demand, and he warned of a supply shock and a period of turbulence.
That is the head of one of the biggest compute buyers telling the biggest memory maker that the order book has some air in it
During a Q&A with Samsung executives, Sam Altman addressed concerns about a bubble in the neocloud sector, saying, “There has been some double counting of demand.”
Asked whether demand for AI, memory chips, and data centers could level off or grow more slowly over the next five to ten years, he replied, “There are ups and downs in the short term.”
“OpenAI and Anthropic are the only two companies really driving substantial growth in demand, so we could see a supply shock and a period of turbulence,” Altman said.
6
Runway just put Praxis 1 out for everyone, and the weights are open under Apache 2.0 with commercial use included.
It is a 27B parameter mixture of experts diffusion transformer that predicts video and robot actions together. The bet is simple. General video is the most scalable robotics data on the planet.
The number that matters is this one. Fine tuned on Google's RT 1, a robot it never saw in training, Praxis 1 hits 80.0% on SimplerEnv. The original RT 1 policy sits at 52.4%.
It already lives inside LeRobot and runs on NVIDIA hardware from H100 all the way down to Jetson Thor.
A video company is now shipping robot brains. Watching the internet turned out to be the cheapest robotics lab ever built
12
Anthropic says Claude Mythos got Claude serving prompts on AMD MI355X GPUs in about 3 days. Almost no human help.
That same job used to eat dozens of engineers and 6 to 12 months of work.
Claude wrote or adapted nearly all of the specialized GPU code on its own. That is the slow painful layer most teams dread touching.
The roadshow also showed revenue up 45 times while compute costs rose only 10 times over six quarters. Fast chip switching is a real part of that math.
Amazon, Google, Nvidia and now AMD all sit in the mix. Compute is starting to look like a commodity Anthropic can shop around for
Anthropic says its Claude Mythos model got Claude processing prompts on AMD's MI355X GPUs in about 3 days, with almost no human help.
That job used to take dozens of engineers 6 to 12 months.
Claude wrote or adapted almost all the specialized GPU code needed to make it work.
1
70
Fable 5.2 cuts input tokens from $10 to $8 per million and output from $50 to $40.
That is 20% lower pricing and up to 30% faster. Lower cost and less waiting make this a real step up.
ANTHROPIC, WHAT THE ACTUAL FUCK???
Claude Fable 5.2 isn't just more capable. It's also faster than its predecessor.
- $8/M input tokens, down from $10
- $40/M output tokens, down from $50
- 20% cheaper
- Up to 30% faster
WHAT IS GOING ON AT ANTHROPIC??
1
1
23
AMD launched a codesign platform with IBM and ORNL QuantumSciCtr for quantum error correction.
Umbra runs under it on Instinct MI355X GPUs and 5th Gen EPYC CPUs beside IBM quantum computers.
Classical silicon is now the control plane for quantum at scale
Quantum progress. Powered by collaboration.
We're excited to announce we're launching a new co-design platform with @IBM and @ORNL's @QuantumSciCtr to advance quantum error correction. At its foundation: Umbra, an HPC system powered by AMD Instinct MI355X GPUs and 5th Gen AMD EPYC CPUs, working alongside IBM quantum computers.
bit.ly/4yXwcHJ
38
Maxinsights closed a $30M seed led by Peak XV.
It delivered 2M+ hours of real world robotics work data, has 50+ robotics and foundation model teams lined up, and projects a $250M annualized run rate for 2026.
Real world data is the scarce layer under robotics AI
Robotics AI is only as good as the real-world data behind it.
That's what makes Maxinsights' latest progress interesting.
→ $30M seed funding led by Peak XV
→ 2M+ hours of real-world work data delivered
→ 50+ robotics and foundation model teams in the pipeline
→ $250M projected annualized run rate for 2026
The race to build smarter robots is accelerating, and the data infrastructure behind it could become one of the biggest opportunities in AI.
1
26
Zipline raised $1.5B at a $21.5B valuation.
That jumps from $7.85B after a $600M raise in January, with US manufacturing next.
Autonomous drone delivery is now priced like a public tech giant
Autonomous drone delivery startup Zipline raised $1.5B at a $21.5B valuation, up from $7.85B after raising $600M in January, and plans to grow US manufacturing (@kelly_cloonan / Wall Street Journal)
(Visit Techmeme dot com for the link and full context!)
38
Agents on Tencent Cloud now spin up 20,000+ PostgreSQL instances a day, up 80x in a year.
14 million sandboxes go up daily so agents can run code and browse the web in isolation.
Cloud demand is turning into agent demand at machine speed
Wow!
Agents are now creating 20,000+ PostgreSQL instances a day on Tencent Cloud.
This is up 80x in a year!
14 million sandboxes are created daily. Agents run isolated environments in these sandboxes to execute code and browse the web.
Tencent's office AI agent has received 9 billion (with a B) task requests in the last 8 months!
14
Dan Ives moved Big Tech capex peak to 2027 after the midterms from his prior 2028 to 2029 window.
10% to 20% of 800 to 1,000 planned US AI data centers face scrap or delay, with Big Tech raises about 25% lower next year
Bullish stays. Selectivity got sharper
DAN IVES SAYS THE MIDTERM RESULTS WILL SLOW THE AI DATA CENTER BUILDOUT
He now expects Big Tech's capex growth to peak in 2027, earlier than his previous 2028 to 2029 view.
- 10% to 20% of the 800 to 1,000 planned US AI data centers could be scrapped or face long delays
- Big Tech's stock and debt raising could be about 25% lower over the next year
- Losers: neoclouds, power suppliers, and some chip and memory companies
- Winners: Apple $AAPL, Meta $META, Microsoft $MSFT, Google $GOOGL, Amazon $AMZN, Palantir $PLTR, Snowflake $SNOW, CrowdStrike $CRWD, Palo Alto Networks $PANW, Nvidia $NVDA
He says his overall bullish AI view hasn't changed. He's just more selective.
1
58
FERC cleared MARA's Long Ridge Energy and Power acquisition.
The plant can sell 485 MW now and is expected to reach 505 MW on more than 1,600 contiguous acres, with closing in the coming weeks.
Grid scale power like this is becoming the scarce asset around AI capacity
The Federal Energy Regulatory Commission (“FERC”) has authorized our acquisition of Long Ridge Energy & Power. Closing is expected in the coming weeks.
The deal will add 1,600+ contiguous acres and a power plant currently authorized to sell 485 MW, expected to reach 505 MW nameplate in 1H 2027. Learn more: ir.mara.com/news-events/pres…
21
AppLovin Q3 revenue hit $2.185B, up 55% YoY versus $2.07B consensus.
EPS printed $4.30 against $4.03 expected. Net income hit $1.44B at a 66% margin, with adjusted EBITDA $1.84B and an 84% margin.
$1.4B free cash flow shows the AI ad stack converting growth into cash
$APP Q3 2026 Earnings — Double Beat
AppLovin delivered a strong Q3 beat on November 4, 2026:
Revenue: $2.185B (+55% YoY) vs. ~$2.07B consensus
Diluted EPS: $4.30 vs. ~$4.03 estimate
Net Income: $1.44B (+72% YoY) | 66% net margin
Adjusted EBITDA: $1.84B (+59% YoY) | 84% margin
Free Cash Flow: $1.4B
43