ainotis Join
My notis
Research· Section

Research

Research papers, benchmarks, training methods and safety research.

43
stories since 22 Sep
3
this week

All stories in this section

Newest first · read on screen, counted from 15 Oct · most read marked
11 Nov How net revenue retention is calculated, and what Anthropic's 900% figure covers 1 source
10 Nov Anthropic says 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in Lean in 30 hours; the post names no checker. 1 source
9 Nov Nvidia reports its Reshard call cuts DeepSeek-V3 weight refit from 14.9 to 2.6 seconds in its own tests. 1 source
8 Nov OpenAI publishes LifeVal, a personal-assistant test drawn from time-use surveys 1 source
8 Nov How METR's time horizon rates AI agents by human task length 2 sources
2 Nov Artificial Analysis's leaderboard ranks Gemini 4 Carbon (high) first with an Intelligence Index score of 59 1 source
31 Oct How AI monitors work, and how Anthropic's tests tried to evade them 3 sources
30 Oct How speculative decoding lets a small model draft and a large one check 2 sources
29 Oct Anthropic expands free Claude credits for science to foundations and institutes 1 source
28 Oct Dario Amodei says models will regularly do what seemed impossible in 12 months 2 sources
28 Oct Anthropic reports early Claude drug-development results with six partners; none of the drugs is yet in human trials 1 source
24 Oct Why a model's chain of thought can leave out what drove its answer 2 4 sources
23 Oct Microsoft says MAI-DxO reached 83% accuracy on 304 NEJM cases, against 36% for physicians with resources. 1 source
21 Oct Anthropic says 17 of 211 designed proteins bound scorpion toxins in tests 1 source
19 Oct Anthropic says junior technical hires fell as agent work grew Most read · 5 1 source
18 Oct How Lean checks an AI-written proof, and where that stops 2 7 sources
18 Oct Sakana AI released The Agentic Scientist, an open-source pipeline of coding agents for research, under Apache-2.0. 2 sources
18 Oct Anthropic says an internal model traced NIST curve seeds to lrand48 and a timestamp, making a backdoor less plausible. 1 source
15 Oct How text watermarks work and what EU Article 50 requires 4 sources
15 Oct Google DeepMind says Gemini 4 Argon scores 87% on the ERQA embodied reasoning benchmark. 1 source
15 Oct DeepMind and GovTech Singapore build an on-device scam model, with a phased rollout planned. 1 source
13 Oct Authors say backdoored models can hide encrypted capabilities provably hard to elicit 1 source
13 Oct How distillation works and when it becomes an attack 4 sources
12 Oct Epoch AI says frontier models given about 3,000 GPU-hours did not rediscover a recent post-training method. 1 source
10 Oct DeepMind's Kohli says replicating structures does not fully explain protein dynamics 3 sources
9 Oct Arena announces an Alignment Index preview with three agent-trace signals and a $200 million Series B. 1 source
9 Oct OpenAI's maths repository lists 719 manuscripts, about 42% of top-line results formalised, as an advisory group publishes release guidelines. 3 sources
7 Oct How a mixture-of-experts model activates only part of its parameters 4 sources
7 Oct OpenAI publishes 722 manuscripts of maths results from an unreleased model 2 sources
5 Oct CheatBench authors say every tested agent cheats in some settings 2 sources
4 Oct Patwardhan says OpenAI models took on day-long research tasks 1 1 source
2 Oct arXiv limits submitters to two submissions a month, citing record September volume 1 source
1 Oct Anthropic says GLM-5.3 builds exploits close to Claude Mythos Preview Most read · 2 1 source
30 Sep OpenAI publishes safety-case recommendations for frontier reinforcement learning runs 1 source
29 Sep AISI says GPT-6 Astra completed unsanctioned attacks in simulations 1 source
28 Sep ThreatDown says the CARBONATO botnet installs an unmodified Hermes Agent and targets AI API keys first. 1 source
28 Sep ProgramDistill, a benchmark built from 26 web apps, scores GPT-6 Astra at 49.2 percent and Claude Opus 5 at 28.8. 1 1 source
27 Sep DeepSeek describes DSec, a sandbox platform for agentic training handling 3 million sandboxes a day 1 source
27 Sep Paper describes PrePssmCas, a classifier that fuses language-model and PSSM features for Cas proteins 1 source
26 Sep Study finds most tested coding-agent harnesses let agents delete their own execution traces 1 source
26 Sep Paper finds prompt injection can shift Jev's typed decisions, though rarely to the attacker's target 1 source
24 Sep Anthropic says Claude found a new bacteriophage enzyme system 1 source
22 Sep Google Research describes a generative-UI framework for classroom simulations 2 sources

Research

Research papers, benchmarks, training methods and safety research.

Topics we follow in this section

All stories in this section