My advice is to not try to get it to one-shot it. Actually I have a pet theory that very stern instruction primes a junior developer role (because that's who gets micromanaged). Just invoke a second process, it's better for the tests not to have the same context anyway
RESEARCH
-
Custom AI Benchmark Created Without User Data
By
–
No user data was used in the making of the custom benchmark, to be clear!
-
Google Releases Gemini 3.1 Pro with Doubled Reasoning Performance
By
–
Today, we’re continuing to push the boundaries of AI with our release of Gemini 3.1 Pro.
— Jeff Dean (@JeffDean) 19 février 2026
This updated model scores 77.1% on ARC-AGI-2, more than double the reasoning performance of its predecessor, Gemini 3 Pro.
Check out the visible improvement in this side-by-side comparison,… pic.twitter.com/8fbHFLeyL8Today, we’re continuing to push the boundaries of AI with our release of Gemini 3.1 Pro. This updated model scores 77.1% on ARC-AGI-2, more than double the reasoning performance of its predecessor, Gemini 3 Pro. Check out the visible improvement in this side-by-side comparison, showing Gemini 3.1 Pro’s crisp animation built with pure code. Read more about today’s 3.1 Pro update: blog.google/innovation-and-a…
-

Gemini 3.1 Pro Launches: Deep Reasoning Model on Poe
By
–
Gemini 3.1 Pro is live on Poe — Google’s newest Gemini model built to solve your hardest challenges: deep reasoning, synthesis, and complex problem solving. We got early access to a private snapshot and co-developed a custom benchmark to help shape this model with our partners at
-

Gemini 3.1 Pro Achieves SOTA Performance on AI Benchmarks
By
–
BREAKING : GEMINI 3.1 PRO HITS 77.1% ON ARC-AGI-2 AND 44.4%. ON HLE. ON PAR WITH OPUS 4.6 ON SWE BENCH VERIFIED. ABSOLUTE SOTA
-

Gemini 3.1 Pro Launch: Advanced Capabilities and SVG Improvements
By
–
Gemini 3.1 Pro has landed! Amazing performance / capabilities across the board. Beyond SOTA, the best are all the things that evals can't measure. E.g. SVG has gotten so much better (see 🧵) blog.google/innovation-and-a…
→ View original post on X — @oriolvinyalsml, 2026-02-19 16:21 UTC
-
Anthropic’s Document Assembly Capabilities Debate
By
–
Without knowing the details, my best guess is that it is just not that good at ppt/excel – a lot of these tasks are not blocked by model smarts but just how well it can put a document together. Anthropic invested in this ahead of others. Useful capability but not load bearing
-
Stochastic Gradient Lattice Random Walk for Bayesian Inference
By
–
Struggling with minibatch noise in Stochastic Gradient Bayesian Inference? Want your chains to naturally run on stochastic hardware?
— Lars Holdijk (@HoldijkLars) 19 février 2026
Introducing — Stochastic Gradient Lattice Random Walk!
New work from @NormalComputing in collab with @zierhjmensch, @adnarim066, @wellingmax, and… pic.twitter.com/Mt7CBwopMaStruggling with minibatch noise in Stochastic Gradient Bayesian Inference? Want your chains to naturally run on stochastic hardware? Introducing — Stochastic Gradient Lattice Random Walk! New work from @NormalComputing in collab with @zierhjmensch, @adnarim066, @wellingmax, and the dream team at normal computing, @Sam_Duffield, @MaxAifer and @ColesThermoAI.
→ View original post on X — @wellingmax, 2026-02-19 16:05 UTC
-

Gemini 3.1 Pro Preview Available on Google AI Studio
By
–

BREAKING : Gemini 3.1 Pro Preview is now available on Google AI Studio! "Our latest SOTA reasoning model with unprecedented depth and nuance, and powerful multimodal understanding and coding capabilities"
-
Gemini 3.1 Pro Preview Introduces Medium Thinking Level for Enhanced Reasoning
By
–
Key changes compared to Gemini 3 Pro Preview Gemini 3.1 Pro Preview adds support for the MEDIUM thinking level, similar to support in Gemini 3 Flash Preview: – Low: Minimizes latency and cost. Best for simple instruction following or chat.
– Medium: Balances reasoning depth and