Engineer @GoogleDeepMind. Prev. Product @ GDM, Founder/CEO @ Scaled Inference, Engineer @Google (Search, Research, X, Brain).
Palo Alto, CA
Joined June 2008
- Tweets7K
- Following1.6K
- Followers4.3K
- Likes49.8K
Olcan retweeted
i've reproduced the results from the MIT blog post today on weight-space programming. but one additional result i have to add: this is a new, step-change emerging capability that's appeared in just the latest set of models.
Olcan retweeted
AI may well be able to write massive software projects “from scratch”, but that almost certainly isn’t the fastest / highest performance / highest security / most token efficient way to do things. Good building blocks matter.
In addition to RL’ing the agents on final task loss, we should try to get them to regularize the building blocks as well. What would the standard library look like that makes the agents most effective?
It is likely at the limit that it turns out to be some alien organization of code, but given the existing human-bias in training, I bet that in the near term it would be lovely code that old programmers would appreciate.
I suspect the same could be said for mathematics, but that is out of my lane.
Olcan 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!
Olcan retweeted
A ‘hello world’ milestone for Project Suncatcher 🚀
After launching our prototype satellite a few weeks ago, we’ve now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth - a haiku!
Olcan retweeted
Less than a year ago, launching Gemini 3 Pro was a big-bang event where the whole team celebrated #1 benchmark numbers.
Today Gemini 4 Carbon hits #1 on the AA Index, yet the feeling inside is distinctly different. We've been dogfooding continuously for weeks, and we've almost adapted to the new intelligence -- taking ambitious code rewrites for granted, leaving loops overnight that launch dozens of experiments to repro, root-cause, and fix mysterious bugs, and using swarms to inspect and fix RL data.
Launches just feel like periodic checkpoints now.
Just two weeks after rolling out Gemini 4 Argon, excited to share a preview of Gemini 4 Carbon - our next frontier model. It has state-of-the-art long horizon coding capabilities driving SOTA performance across the cybersecurity defense lifecycle including detection, validation, and remediation. It also ranks #1 overall on the recently updated v5.0 of Artificial Analysis Intelligence Index at 59.
We're starting by rolling out to a subset of cyber defenders through our Fairwind Program this week and soon expanding access further. Read more at: blog.google/innovation-and-a…
Olcan 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!
Olcan retweeted
People have expected that RSI would magnify small leads and mean one lab runs away with it all. That may happen, but the frontier AI race also has a built in self-stabilizing dynamic that pushes the other way.
If you build a spectacular model, you sell out and may have to canibalization all your training and research compute to stay afloat and keep your customers from rioting, as well as needing to direct a lot of your efforts to inference efficiency on the existing model. So your next model’s quality may suffer a lot.
If you release a really bad model, no one uses it, and you have a ton of unneeded inference compute, and can scale faster and experiment more for your next model runs, which can mean a much better model.
Or, in meme form: Big compute builds strong models. Strong models lead to little compute. Little compute leads to weak models. Weak models lead to big compute.
Olcan retweeted
I’m so happy to be able to share some of our work with you. Our model Ampersand, we trained to help people come closer to their dreams.
I love the humans& team and am so glad I can share this moment with them and you.
No one changes the world alone
Introducing Ampersand - the first model built for us. Ampersand is a collaborator trained to help people achieve their individual and collective goals. In multi-month simulations, users collaborating with Ampersand accomplished more than with Fable and other frontier models
Available to everyone through our API
Olcan retweeted
Astra dethroned
Gemini 4 Argon is the best "vision" model
#1 for detection:
- tight, precise boxes
- dense scenes
#1 for text extraction:
- small text
- handwritten text
- complex tables
- rotated text
4× cheaper than Astra at high
Olcan retweeted
Argon is so good at robotics that our whole team is having an existential crisis right now, enjoy!
developers.googleblog.com/ge…
Olcan retweeted
Gemini 4 Argon (gemini-4-argon) is now live across Google AI Studio, the Gemini API, Google Antigravity, Gemini Enterprise, and the Gemini app!
Gemini 4 Argon is awesome:
1. Unified 1M Context & 1M Output (Long Decode): Input, thinking, and output now share a single 1M-token window—lifting the old 64K output cap by 16x. With background=True on the Interactions API, our servers handle multi-hop continuation and checkpoint recovery automatically so the model can think for tens of minutes and emit complete multi-file repo migrations or formal proofs in a single pass.
2. Tunable Reasoning Depth: Control compute per turn via thinking_level ("low" | "medium" default | "high") plus native agentic video processing (media_processing="agentic").
3. Real-World Engineering: 77.9% on DeepSWE v1.1, 91.9% on Vibe Code Bench, and 68.9% on the Vals Index. Internally, we used Argon to migrate 800K+ lines of Fuchsia Zircon C/C++ kernel code to memory-safe Rust and replace 32K lines of hand-written SIMD in libgav1 for a 2.7x speedup.
4. Token Economics: Launching at an introductory rate of $2 / 1M input and $10 / 1M output tokens through Nov 30, with 95% off cached inputs ($0.10 / 1M) via automatic implicit caching.
Olcan retweeted
(1/7) I'd like to provide some additional context for the remarks that can be found in this interview.
(2/7) My discomfort with the attention being paid to existential risk---and with the ongoing dominance of the public discussion of such risks by a few so-called thought leaders---is that there's an elephant in the room that is being missed. That elephant is the effect of all of this discussion on young people, and indeed the effect---beginning with social media a decade ago---of information technology on young lives broadly speaking.
(3/7) I recently read an interview by Jill Lepore in which she reports on discussions that she had recently with a group of college students: "They had just three things to say. One, social media completely destroyed their childhoods. Two, AI is destroying their college education and their likely prospects for employment. And three, what the hell are we all doing?" To add a fourth bullet that I've heard in similar discussions: "I'm not going to have children, because the existential risk is just too high."
(4/7) Please contemplate such a statement, those of you who think that it's OK that the public, widely diffused, discussion coming from seemingly authoritative figures is so adamant about existential risk.
(5/7) Let me add that I have little interest in a debate about what "superintelligence" might mean, or whether we're currently seeing the emergence of superintelligence. For what it's worth, I would certainly agree that the solution of extremely challenging math problems (for example) exhibits a very impressive form of intelligence, although I would want to argue that such intelligence is impossible to dissociate from the collective of humans that provided data to the machine learning system (which itself has become a kind of collective).
(6/7) Such philosophical debates are entertaining and even important, but when they then head in the direction of "when such superintelligence emerges, there's a good chance that it will want to kill all humans", I say "science fiction". Again, I don't think that it's bad to consider such thoughts, but I strongly believe that it's bad for society, and especially young people, for such thoughts, unaccompanied by any serious real-world analysis, to become a dominant form of discourse of the era.
(7/7) We are damaging human lives, and the consequences of that damage will no doubt be a major theme of the next several decades.
Olcan retweeted
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.
- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all.
- Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.
- Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.
- And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.
The funnel is driven by a combination of factors:
1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel.
2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel.
3. Access. Free-tier, paid-tier, internal.
So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.
Judging by my tl there is a growing gap in understanding of AI capability.
The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code.
But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along.
So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions.
TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned (?) "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.
Olcan retweeted
AI is a tool. Like aviation or autonomous driving (Waymo), our job as builders is to build it at full pace while holding ourselves accountable for safe rollouts in the real world.
Crying wolf and demanding we halt only guarantees less accountable actors deploy it first. Anthropomorphizing AI as a "conscious entity" lets builders deflect responsibility.
I reject both.
We need to build AI, at pace, and with full accountability.
Olcan retweeted
Reading all those recent beautiful essays by Terry Tao, Kevin Buzzard and @wtgowers, I could not resist, but write up my deeply personal perspective on how I perceive and see things:
docs.google.com/document/d/e…