I understand technology better than you. Really.
Joined March 2009
- Tweets7K
- Following365
- Followers544
- Likes3.7K
Stefan Norberg retweeted
Introducing Gemini Decision 1, our new frontier decision model in the Gemini API.
It is $0.10 per 1M input tokens with no additional cost. With support for image, audio, video, and PDF natively, 256K context window, & 1K questions.
You can try it for free in the API!
Stefan Norberg retweeted
I’m very pleased to introduce AlphaGenome 2, @GoogleDeepMind’s next-generation genomics model that deciphers how DNA changes affect cells.
Cells are the basic units of life and key to our understanding of biology. It is an ongoing challenge to predict how DNA changes affect the functioning of different cells, such as for a heart muscle cell or an immune cell fighting an infection, and to uncover cellular drivers of disease.
To make progress on this challenge, we needed to design the architecture of AlphaGenome 2 so its predictions were conditional on cell types and states. This new design enables AlphaGenome 2 to predict variant effects across hundreds of human cell types and adapt to unseen cell and disease states without retraining.
When evaluated on the TenK10K immune cell dataset, AlphaGenome 2 demonstrated a major leap in accuracy over its predecessor. It reliably identified nearly twice as many causal variants (30% vs. 16%) and directional gene effects (43% vs. 23%) compared to the original AlphaGenome. This step-change improvement in precision helps researchers pinpoint previously overlooked disease-relevant variants with high confidence.
Using analysis on retrospective case studies, we were able to show AG2’s real world utility in disease understanding. A landmark study had previously shown that a variant in a distant regulatory switch drives irritable bowel disease by overactivating the ETS2 gene in gut macrophages. When used to make predictions across various immune cell types in the gut, AlphaGenome 2 singled out the exact causal variant and showed that it selectively boosts the ETS2 gene in gut macrophages, leaving other cell types and genes largely unaffected.
I would like to thank my colleagues Adam Gayoso @adamgayoso, Žiga Avsec @Avsecz, and our entire genomics team for their incredible contributions and dedication to bringing this model to life.
We are making AlphaGenome 2 available for non-commercial academic research via the AlphaGenome API, and soon we will release AlphaGenome 2 Skills in Google Antigravity. I’m incredibly excited to see how scientists around the world use this model to make progress on biological understanding and treating disease.
Read more on our blog here:
deepmind.google/blog/alphage…
Stefan Norberg retweeted
1/2 Gemini 4 Carbon land today. and (general) vision perception is almost solved in the agentic era, we need more
Humanity's Sixth Sense by @scale_AI just release at Oct. it said "far from saturated despite high human performance". but Gemini+Antigravity easily saturate it 🤯
Stefan Norberg retweeted
Across our projects @GoogleDeepMind, we are witnessing a genuine phase transition in the pace of scientific discovery. You will see a glimpse of this in our releases in the next few weeks.
Stefan Norberg retweeted
This is pretty much how things have gone...
ANTHROPIC: "AI is extremely perilous and should be carefully regulated"
CALIFORNIA GOV: "Ok, we will regulate you."
ANTHROPIC: "No! Good regulation is really hard to get right. We know so little about what to do. Instead we should let companies carefully iterate over time."
CALIFORNIA GOV: "Oh ok. How about we do a thing where you get to write what the rules are but we will hold you to them? Then you can make your own commitments as you think is best but they will be true commitments!"
ANTHROPIC: "Nah, we'd rather make fake commitments."
CALIFORNIA GOV: ...
If Anthropic wants me to think these steps are real/credible, they should put it in their SB 53/RAISE Act Frontier Compliance Framework.
State law gave Anthropic and OpenAI a mechanism for making credible statements about their behavior, and they mostly choose not to use it on the risks they are most worried about (neither company includes anything in their compliance frameworks about internal deployment RSI risk, though Anthropic says maybe at some point in the future they will. Now seems like as good a time as any???!!!???)
Stefan Norberg retweeted
Sharing early evals on Gemini 4 Carbon. Strongest model that we've seen so far for long-form research and answer quality, with a new SoTA on the Wide-And-Deep-Research (WANDR) benchmark. Coming soon to all Perplexity Pro and Max users inside Computer, once we have sufficient token capacity. Congrats @GoogleDeepMind!
Stefan Norberg retweeted
"So Google is spending record sums to build a product that earns less than the one it replaces."
This is all but silly. If they don't do it then someone else will do it.
I've seen the cannibalisation argument so many times.
Reality is this:
People at company X push back on a new X product that would cannibalize sales of X flagship product. Then company Y ships a similar product and X loses it all.
What X should do is to cannibalize itself as fast as it can, faster than the Ys.
A year ago, Microsoft, Alphabet and Meta generated $60.7 billion of free cash flow in a single quarter. This summer, $7.4 billion.
They didn't sell less. Revenue grew 19% at Microsoft, 25% at Alphabet and 26% at Meta. What changed is the bill: $125.6 billion in three months for data centers, chips and servers, double what they spent a year earlier. Meta put 54 cents of every dollar of revenue into capex. Alphabet spent $9 billion more than it took in, its second negative quarter in a row, and this morning its stock fell 12%. If it closes down more than 9.77%, it will be its worst day after earnings since going public in 2004.
The detail that worries me most is in Google's search numbers. Search ads grew 12% and still missed expectations. Philipp Schindler, Alphabet's chief business officer, said part of the reason is that AI Mode makes less money per search. So Google is spending record sums to build a product that earns less than the one it replaces.
For twenty years these were the best businesses ever invented. You wrote the code once, sold it to billions of people, and almost every new dollar of revenue fell to the bottom line. AI doesn't work like that. Every answer runs on a chip someone had to buy and on electricity someone has to pay for. Big Tech is turning into something closer to a utility: huge upfront costs, long paybacks, and profits that depend on keeping the machines full.
That can still be a great business. But it is a different business, and investors are only starting to price it that way.
The old moat was code. The new one is cash.
Stefan Norberg retweeted
My key takeaways from the $GOOGL (Alphabet) Q3 2026 earnings call.
1. Search slowed while query growth hit a five-year high. Search and other ads grew 12% to $63.3B, down from Q2. The deceleration has four parts. FX was the biggest drag. The EU ended its de minimis exemption (duty-free treatment for cheap parcels), which hit APAC retail advertisers. World Cup media budgets landed in June instead of July. And AI Mode queries grew fast on a surface with "lower monetization" than classic Search. The first three wash out over time. The fourth is structural: more usage on a surface that earns less per query.
2. The EPS miss is mostly a SpaceX mark. Revenue grew 25% to $127.6B. Operating income grew 30% to $40.6B, a 31.8% margin. EPS came in at $2.12 against $3.02 expected because OI&E (other income and expense, the non-operating line where investment gains and losses land) swung to negative $8.2B. Most of that was an unrealized loss on the SpaceX stake after its share price fell. My rough math: after tax, that mark costs around $0.55 a share, so it explains more than half the $0.90 gap but not all of it.
3. Q4 Search is already guided softer. Lapping last year's Search acceleration cost about 3 points of growth in Q3 and will cost a couple more in Q4. FX gets slightly worse at current spot rates, mostly in Search and YouTube. The rest of Services won't make up the difference: YouTube ads grew only 9% to $11.2B, and Network grew 1% to $7.5B.
4. The AI Mode ad fix is rolling out but not proven. Ads in AI Mode go international for the first time, and multi-turn ads (ads that stay in a back-and-forth conversation) launch in the U.S. The evidence offered: inside AI Mode, sponsored and organic product listings get similar click-through rates. AI Mode queries run about 3x longer than classic searches, and most are incremental, meaning they would not have happened otherwise. What's missing is any number on how far AI Mode revenue per query trails classic Search. Agentic commerce (AI agents that shop and buy for users) is called additive, and consumer agents arrive in the Gemini app later this year. No revenue is attached to it yet.
5. Cloud doubled. Revenue grew 103% to $30.7B. Operating income was $11.1B, more than triple last year. Margin was 36.2% versus 23.7% a year ago. Backlog (signed contracts not yet booked as revenue) reached $561B, and just over 55% of it, about $310B, converts within 24 months. Two caveats. TPU system sales contributed to the growth but were not sized, and Cloud growth excluding TPUs was described as very strong with no number attached. Also, the rented third-party compute capacity came online only at the very end of Q3, so its margin drag lands in Q4.
6. TPU economics stay private. A TPU is Google's in-house AI chip, sold here as full racks with networking, storage and software. One analyst framed a TPU system at roughly $35B per gigawatt and asked how much of that becomes revenue and at what margin. Neither got an answer, because product-level financials are not disclosed. The cash cycle is clear. A signed contract shows up in backlog, Google builds inventory, and cash arrives on delivery. Inventory is only built against committed contracts. There are nearly 100 TPU customers, including Anthropic (directly and through Broadcom) and Apple, which is scaling pretraining on TPUs. New distribution comes through Oracle and Crux AI, a joint venture with Blackstone. The meaningful revenue ramp is 2027.
7. Free cash flow went negative. Q3 FCF was negative $9.1B on $51.4B of capex. Operating cash flow of $42.3B also absorbed the TPU inventory build and a $5.2B payment on a 2018 EC fine. Trailing twelve-month FCF is $19.7B. The 2026 capex guide is unchanged at $195B to $205B, and 2027 rises significantly, with numbers on the Q4 call. On top of that sit $924B in future purchase commitments and $152B in leases that haven't started and other obligations. Those are contractual maximums, and some have volume flexibility.
8. The return math is stated, not shown. Return on invested capital is said to stay meaningfully above the cost of capital, with no figure. Compute allocation is described as a long-range plan run on ROIC discipline, with no ranking of who gets capacity first. Cloud says it could serve a lot more demand with more capacity. About 65% of Q3 infrastructure spend went to servers and networking depreciated over six years, while TPU 8t claims up to 2.7x the training performance of the prior generation. My view: six years is a long life for hardware that loses that much ground each generation. Shorter lives would mean higher depreciation on top of the capex.
9. First-party model APIs process nearly 32B tokens (units of text an AI model reads or writes) per minute, up 40% QoQ and more than 4x YoY. In its first week on developer platforms, Gemini 4 Argon ramped 3.5x faster than Gemini 3, even with capacity limits. Customers who used Gemini Enterprise in Q3 2025 have more than quadrupled their spend. The Gemini app is at 1B users. This is the clearest evidence that the capex has real buyers.
Bottom line: The operating business beat. Revenue grew 25%, operating income grew 30%, Cloud doubled at a 36% margin, and the headline EPS miss is mostly a paper loss on SpaceX. The stock still fell about 10% after hours, from $322.22 to $288.35, because the market is now pricing cash, not growth. Quarterly FCF is negative, capex runs near $200B and rises again in 2027, commitments total about $1T, and the biggest profit pool is slowing as usage shifts to a surface that earns less per query. Management described every link between the spend and the payoff and did not put any number on them.
Stefan Norberg retweeted
Google’s Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5, and second only to Claude Opus 5.5 on the Artificial Analysis Coding Agent Index
Gemini 4 Carbon scores 59 on Intelligence Index v5, driven by leading performance in complex reasoning, knowledge, and scientific tasks. It also has large gains in coding and knowledge work, particularly when compared to other Gemini models, but it is not leading in all these areas.
As with launch of Gemini 4 Argon, Gemini 4 Carbon is being rolled out to selected users and is not yet publicly available. Google is offering a 50% launch discount, with no confirmed end date. All our charts and figures use standard pricing for both Gemini 4 Carbon and Argon.
Key results for Gemini 4 Carbon with high reasoning:
➤ Gemini 4 Carbon leads the Intelligence Index v5, with the #1 score in 4 of the 9 Index evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, AutomationBench-AA. Carbon makes substantial gains elsewhere but is not leading in coding and knowledge work
➤ Behind only Claude Opus 5.5 on coding agents: With Antigravity CLI, Gemini 4 Carbon (high) scores 59 only 3 points behind Claude Code with Claude Opus 5.5 (max, 62) on the Artificial Analysis Coding Agent Index. Carbon ranks second on both Terminal-Bench 4.0 (64%) and FrontierCode v1.1 (51%), behind Claude Opus 5.5 (max) at 56%. Google has historically underperformed in this area and so this improvement stands out
➤ Improved agentic knowledge work, but not yet leading: On AA-Briefcase, Gemini 4 Carbon (high) scores 1609 Elo, a 126-point improvement over Gemini 4 Argon (high, 1483), but still behind models including Grok 4.7 (xhigh, 1638), Muse Wonder (max, 1659) and Kimi K3.1 Pro (1678). The largest gain comes from Presentation, up 293 Elo to 1598, alongside a 120-point improvement in Analytical Quality to 1671
➤ Lower Cost per Task than Claude Opus 5.5 for comparable overall intelligence: Gemini 4 Carbon costs $7.76 per Intelligence Index task, ~23% less than Claude Opus 5.5 (max, $10.04) but ~1.2x GPT-6 Astra (max, $6.25). It uses 123k output tokens per task, 29% fewer than Claude Opus 5.5 (max, 173k). At the same $4/$20 standard pricing, its Cost per Task is in line with Gemini 4 Argon ($7.68), while scoring 9 points higher on the Intelligence Index
➤ Leading AA-Omniscience v1.1 score, with low hallucination: Gemini 4 Carbon scores 54 on our recently updated AA-Omniscience v1.1, 6 points ahead of Claude Opus 5.5 (max, 48) and 8 points ahead of Gemini 4 Argon (46). Accuracy (% correct) rises 10 points over Argon to 58%, while its hallucination rate rises from 4% to 11%. This is still the second lowest of any model, behind only Argon, compared with 50% for both Claude Opus 5.5 (max) and GPT-6 Astra (max)
Key model details:
➤ Context Window: 1M tokens
➤ Multimodality: Text, image, video, and speech input, with text output
➤ Pricing: $4/$20 per 1M input/output tokens, the same as Gemini 4 Argon’s standard pricing, with a 50% discount. Cached input tokens receive a 95% discount ($0.20 per 1M)
Congratulations to the @GoogleDeepMind team!
Stefan Norberg retweeted
I did say we'd be releasing our models periodically...introducing Gemini 4 Carbon, our next frontier model.
Following Argon's release just two weeks ago, Carbon shows the pace at which we’re pushing the frontier, it ranks #1 on the recently updated v5.0 of @ArtificialAnlys Intelligence Index at 59. It also leads across key cybersecurity evaluations scoring 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching.
We’re seeing significant improvements in long-horizon tasks even compared to Argon, and we’ve strengthened frontier safeguards.
As we did with Argon, we’re going to roll it out gradually + responsibly. Some trusted cyber defenders will get access starting this week. We'll be expanding it to more partners through our Fairwind Program, ahead of wider availability soon, so hang tight!
Stefan Norberg retweeted
Glad to see the improvement purely through switching to the Antigravity harness. Always more work to do but working hard to make sure it’s the best harness across all frontier models!
Gemini 4 Argon is underrated 😱.
When auditing the traces, we found that Pi Harness just ends the run whenever it hits a fallback (e.g. for security task) in 20% of the runs. We rerun Argon with Antigravity Harness, and now it beats Astra and sits right behind Opus 5.5. Congrats to the team @GoogleDeepMind and looking forward to the new ckpt.
Interesting interview with Google VP of Engineering Bill Jia regarding $GOOGL CAPEX investments around AI.
youtube.com/watch?v=w253sKcS…
87
Stefan Norberg retweeted
¡Hola, Region of Madrid! 🇪🇸 We’re thrilled to announce that we’re coming to Spain, beginning with Madrid. This will be our first region in Spain and second in the European Union. We plan to launch fully autonomous ride-hailing in Madrid in 2027.
Sign up for future updates: waymo.com/waymo-in-spain
¡Hola, Comunidad de Madrid! 🇪🇸 Nos emociona anunciar que llegamos a España, empezando por Madrid. Esta será nuestra primera región en España y la segunda en la Unión Europea. Planeamos lanzar nuestro servicio de ride-hailing totalmente autónomo en la Comunidad en 2027.
Stefan Norberg retweeted
We’re bringing the magic of Waymo to Toronto! 🇨🇦
In the coming days, you’ll begin seeing Waymo’s all-electric vehicles driving through Toronto’s streets as we lay the groundwork for our ride-hailing service.
Our goal is to begin serving riders in 2028.
Sign up for future updates: waymo.com/waymo-in-canada
Stefan Norberg retweeted
BREAKING NEWS🚨: Google has submitted its FIRST TPUv7 FP8 vLLM results to our Agentic Inference Benchmark and has demonstrated 1.67x better perf per dollar than even GB300 NVL72 FP4. 🚀
It is amazing to see Google go full steam ahead with externalizing the power of TPUs outside of Google.👇️ (1/4)🧵
Stefan Norberg retweeted
“Any year now…” $TSLA
Stefan Norberg retweeted
Using raw lidar data is exactly the same as showing raw RGB camera data for every pixel and saying "tell me what you see here". Let's not deceive the public. Here's a 100% lidar-only scene. Let the people decide for themselves what they see.
Stefan Norberg retweeted
Today, our team is sharing a preprint on our new system for generative de-novo enzyme design, AlphaProtein Novo.
At @GoogleDeepMind, we have long focused on the building blocks of life (proteins). We developed AlphaFold to predict structure, trained language models to understand function, and built AlphaProteo to design de-novo binders. But proteins play many important roles in nature, such as enzymes that catalyze chemical reactions.
Designing entirely new enzymes from scratch to catalyze custom reactions has long been a holy grail for scientists. Our team has shown how AlphaProtein Novo can produce new-to-nature enzymes with state-of-the-art activity on 2 benchmark reactions. We’ve also designed custom enzymes to synthesize the pharmaceutical motif piperidine, and others to degrade the environmental plasticizer toxin DEHP.
While there is much more work to do in this direction, we’re excited to have demonstrated the ability to design functional proteins that go beyond what is known in nature. Huge congratulations to @jueseph, @ZvxyWu, Josh Abramson, the @GoogleDeepMind protein design team, and our collaborators in the @francesarnold & Liu Group! 🎉
🔗 Preprints and code:
biorxiv.org/content/10.64898…
biorxiv.org/content/10.64898…
github.com/google-deepmind/a…
