Twitxy Saved

Silas Alberti

Verified account Cognition

@silasalberti

chief research officer & founding team @cognition | prev: ai phd @stanford, jane street, deepmind
San Francisco sil.as Joined February 2012
630 Following 14K Followers 854 Posts

Latest posts 5

Pinned post

Silas Alberti Verified account @silasalberti
Excited to announce today at CoreWeave Fully Connected that Cognition becomes the first Vera Rubin customer. We’re very grateful for the partnership with @CoreWeave and @nvidia!
Cognition @cognition Sep 30
Cognition is the first customer running @NVIDIA Vera Rubin, powered by @CoreWeave!

Hopper in 2022. Blackwell in 2024. Now Vera Rubin:

On SWE-2 inference, the new chips deliver about 4.8x more token throughput than GB200 at the same decode speed.

More compute, better agents!
View quoted post on X
7 replies 6 reposts 148 likes 13K views View on X
Silas Alberti Verified account @silasalberti
We started this project only 6 weeks ago. It's the first model we trained with a midtraining stage, combined with many improvements in RL. The biggest relative gains compared to a base model we've achieved so far.

One goal we've been thinking about is how to provide the benefits of open weights without the cyber safety risks. The ability to run models on our own infra or post-train them was crucial when we got started, so we want to give back to the community. On the other hand, increased cyber capabilities in the hands of bad actors are becoming a major risk, e.g. the recent Binance hack by hacker group in North Korea - the largest heist in human history.

We're working on a program that allows us to hopefully get the best of both worlds (with some compromise). Under this program, we would be able to share the weights securely with enterprises, industry partners, and research groups - without risking abuse by bad actors. We're still working on the details and will share more in a few weeks, especially as we release even stronger models.
Cognition @cognition Nov 4
Introducing SWE-2 Flash, our new fast & efficient coding model.

On leading evals, it outperforms all American open-weight models, even ones 3x larger.

To balance transparency and cyber risks, we are building a program to securely share the weights with partners & researchers.
View quoted post on X
18 replies 19 reposts 219 likes 14K views View on X
Silas Alberti Verified account @silasalberti
The Mercor story is super inspiring and it’s been incredible as a friend to witness them grow from 3 guys in an apartment to one of the leading AI companies. When they recently started switching to Devin, I knew our product team cooked because Mercor has always been at the cutting edge of coding tools
Cognition @cognition Oct 28
Founded by three 22-year-olds in 2023, @mercor crossed $2B ARR in June, only 4 months after hitting $1B.

Today, millions of experts on the platform bring their knowledge of medicine, law, engineering, and beyond to frontier AI - helping models take on increasingly complex and consequential work in the real world.

Mercor's engineers move at extraordinary speed.

They give Devin long-running tasks before bed, during meetings, or on their way to the gym. They come back to tested features and PRs ready to merge.
View quoted post on X
2 replies 7 reposts 85 likes 8.2K views View on X
Silas Alberti Verified account @silasalberti
Looking forward to going back to Stanford on Thursday for a guest lecture at CS 146S. Last year I gave "an opinionated guide about AI coding in 2025" and it was all about IDEs & agents. It seems like we've gone through multiple iterations of form factors since. This year I'll give a refreshed guide for 2026 and it looks very different.
Mihail Eric Verified account @mihail_eric Oct 27
@silasalberti, the Chief Research Officer at @cognition (the team behind Devin, Windsurf), is guest lecturing this week in my Stanford course 𝗧𝗵𝗲 𝗠𝗼𝗱𝗲𝗿𝗻 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿.

Cognition is on a tear, with its valuation rising from $26B to $48B in four months. Silas leads the research behind its coding models, including SWE-2, the first one they trained with effort levels.

Week 6 is all about agentic code review, and I'm excited to hear how a team training frontier coding models thinks about it.

𝗧𝗵𝘂𝗿𝘀𝗱𝗮𝘆 𝗮𝘁 𝗦𝘁𝗮𝗻𝗳𝗼𝗿𝗱 (𝗦𝗧𝗟𝗖 𝟭𝟭𝟭) · 𝟱:𝟯𝟬-𝟲:𝟮𝟬𝗽𝗺. Join us!
View quoted post on X
7 replies 2 reposts 75 likes 3.6K views View on X
Silas Alberti Verified account @silasalberti
Excited to see FrontierCode 2.0 be a part of the new Artificial Analysis Index v5!
Artificial Analysis Verified account @ArtificialAnlys Oct 28
Announcing Artificial Analysis Intelligence Index v5, featuring 2 new benchmarks and 3 upgrades. For the first time, the majority of the Index is built on private datasets

The Artificial Analysis Intelligence Index is our composite measure of model capabilities across tasks reflecting real-world usage today. v5 is our biggest upgrade since v4 launched in January. With v5, 60% of the Index is now based on evaluations with private test sets or held-out answers.

Alongside Intelligence Index v5, we’re launching the Artificial Analysis Progress Index v1, which tracks model capability over time on one continuous scale.

New to the Intelligence Index:

➤ FrontierCode 2.0: Cognition's newly-launched private benchmark of real-world coding tasks, graded on whether a repository maintainer would merge the agent's change

➤ Terminal-Bench-Science: 70 agentic scientific research tasks, run in a terminal environment

Upgraded:

➤ HLE-Diamond (replacing Humanity's Last Exam): a refined 1,000-question version of Humanity's Last Exam, split evenly between reasoning and knowledge questions

➤ AA-Omniscience v1.1: An updated dataset with saturated and public set questions removed, new questions added, and ambiguous or erroneous questions revised

➤ CritPt v2: a private, 40-question hard subset of CritPt v2, revised by the CritPt authors and run with a code-execution tool

Staying in the Index: AA-Briefcase, AutomationBench-AA, GDP.pdf and Terminal-Bench 4.0

Removed from the Index: GDPval-AA v2.1, SciCode, and AA-LCR v1.1

Key results:

➤ Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max, 57) and GPT-6.1 Sol (max, 55): Anthropic and OpenAI hold the top five models. Gemini 4 Argon (high) is next at 50

➤ The open weights gap widens from 3 points to 12: Kimi K3.1 Pro (weights not yet released) is the top open weights model at 46. Among released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) lead at 32 - 26 points behind the leader and below Claude Haiku 5.5 (max, 36)

➤ Cost per task varies by more than 13x among the leading models: of the 10 highest-scoring models shown, GPT-6.1 Terra (max) is the cheapest at $0.96 per Intelligence Index task and Claude Fable 5.1 (max) the most expensive at $13.04
View quoted post on X