01
AI ResearchinclusionAI's Duo v1 Answers Video and Speech With Video
inclusionAI released Duo v1, an Apache-2.0 model that answers a user's video and speech with generated video and audio. Here's what it does and doesn't do.
Frontier models, chips and the science of intelligence
01
AI ResearchinclusionAI released Duo v1, an Apache-2.0 model that answers a user's video and speech with generated video and audio. Here's what it does and doesn't do.
02
AI ResearchNVIDIA Research's VFlash drafts whole visual primitives like boxes and trajectories in one pass, raising robot task success to 70.2% while cutting latency.
03
AI ResearchGoogle's Gemini 4 Hydrogen claims 2.5x faster tasks than Gemini 3.8 Flash, with four thinking levels and $1.50 per million input tokens for agent work.
04
AI ResearchMIT researchers train climate emulators on synthetic emissions scenarios that look unrealistic, cutting reliance on expensive full-scale climate model runs.
05
AI ResearchGoogle DeepMind's AlphaGenome 2 predicts how DNA variants act in specific cell types, roughly doubling causal-variant recall versus its predecessor model.
06
AI ResearchGoogle DeepMind's agents produced about 6.5 million lines of Lean proofs verifying Signal's libsignal code for forward secrecy and post-compromise security.
07
AI ResearchMETR says Anthropic's model progress has roughly doubled, and METR puts AI at about 80% of the speedup. Here's what the audit found and where doubts remain.
08
AI ResearchGoogle's new TPU 8tx targets confidential inference for regulated enterprises, while the dual-chip 8i Max doubles TPU 8i compute for long-context agents.
09
AI ResearchThinking Machines releases Inkling 2 Small, a 594B-parameter model with closed weights for now, promising Apache 2.0 release after external safety testing.
10
AI ResearchOpenAI's GPT-6 Cyber, rated Critical, reports 948 Linux kernel flaws and 81 Firefox bugs, with access limited to vetted customers under strict controls.
11
AI ResearchMIT CSAIL researchers report GPT-6 Astra edited an open Qwen model's weights directly, lifting its MATH-500 score from 7.9% to 48.1% in one experiment.
12
AI ResearchMistral's Timestral family of four forecasting models, from 5M to 209M parameters, claims state-of-the-art wins on GIFT-Eval, TIME and fev-bench benchmarks.
13
AI ResearchAi2 finds LLM judges of research-idea novelty are unstable, with a single prompt change cutting one model's accuracy from 93.5% to 40.9% on the same pairs.
14
AI ResearchA new benchmark paper shows the top coding agent passes nearly 70% of acceptance criteria on real user tasks but fully completes only 26.4% of the tasks.
15
AI ResearchOpenAI has detailed a four-layer system to detect misaligned frontier AI agents, saying a replay of its Hugging Face incident raised an alert 30 hours early.
16
AI ResearchOpenAI claims its models disproved the Exponential Time Hypothesis and proved local smoothing in all dimensions, with Lean proofs. Experts are checking it.
17
AI ResearchOpenAI's GPT-6.1 Terra costs $1/$5 per million tokens, more than doubles Terminal-Bench scores and cuts hallucinations, but its system card flags agent risks.
18
AI ResearchDeepSeek's V4.1-Flash-1017 open-weight update lifts DeepSWE to 77% and Terminal-Bench to 42.3 with no price change, plus new Search and Fetch APIs.
19
AI ResearchMeta's Muse Spark Open is an 816B-parameter, 1M-token open-weight model for coding and agents that can run locally on two Apple M5 Ultra Mac Studios.
20
AI ResearchMeta Superintelligence Labs' Muse Wonder and Wonder Full reach frontier benchmark scores with 1M-token context, but ship with no open weights at $4/$20 per million.
21
AI ResearchAnthropic's Claude Haiku 5.5 costs $0.10 per million input tokens, a 90% cut, while posting big gains on computer-use and coding benchmarks over Haiku 4.5.
22
AI ResearchMistral AI unveils Large 4, a 1-trillion-parameter open-weight MoE model with 49B active parameters, strong cyber-defence scores and weights due Oct 27.
23
AI ResearchOpenAI's GPT-6.1 Sol claims near-Astra coding and agent skills at $2/$10 per million tokens, while the planned GPT-6.1 Astra was shelved over safety concerns.
24
AI ResearchGoogle unveils Gemini 4 Argon, its first Gemini 4 frontier model, with 1M-token output and top coding scores, rolling out first to vetted cyber defenders.
25
AI ResearchDeepSeek released open-source tools including TileLang, DeepGEMM and FlashMLA for Huawei's Ascend 950 chips, a direct challenge to Nvidia's CUDA software lead.
26
AI ResearchAnthropic's Claude Sonnet 5.5 nearly matches Opus 5.5 on key benchmarks at $2/$10 per million tokens, but independent tests show max-effort runs can cost more.
27
AI ResearchAlibaba's Qwen-Image-2.1 packs image generation and editing into 7B parameters, runs on one consumer GPU, and claims wins over closed rivals. The license has a catch.
28
AI ResearchAnthropic says Claude computed the nine-loop six-particle amplitude in N=4 super Yang-Mills, one loop past the 2023 record, for about $1,000 to $2,000.
29
AI ResearchGoogle's Project Suncatcher will launch four TPUs into orbit on October 1 aboard SpaceX Transporter-18, the first real test of whether AI chips can run in space.
30
AI ResearchAnthropic's new biology lab says 950 Claude agents screened 200,000 reverse transcriptases in 21 hours and found ART, a CRISPR-like phage enzyme system.
31
AI ResearchOpenAI released GPT-6 Sol and Luna with API prices cut by half or more, pushing the frontier on cost per task rather than on benchmark ceilings.
32
AI ResearchStepFun's Step 5 Preview pairs a 600B sparse MoE with a 1M-token context and $1 input pricing, with open weights promised for 15 October.
33
AI ResearchAnthropic's new R&D Automation Index puts Claude at the lead of 26% of its internal AI research work, up from under 1% in February 2026.
34
AI ResearchDeepSeek's MIT-licensed V4.1-Flash packs 552B parameters, a 1M-token window and a KV cache footprint cut to a quarter of its predecessor's.
35
AI ResearchOpenAI says Astra is the first model to meet the Critical cybersecurity threshold in its Preparedness Framework, triggering gated release and new safeguards.
36
AI ResearchNvidia put Vera Rubin NVL72 with Groq 3 LPX into full production, claiming 3,400 output tokens per second on 100,000-token agentic workloads.
37
AI ResearchMoonshot AI's Kimi K3 ships full open weights at 2.8 trillion parameters, the largest open model yet, intensifying the open-versus-closed frontier race.
38
AI ResearchOpenAI says its unreleased Astra model resolved ten decade-old math problems for about $2,000, releasing machine-checkable Lean proofs for independent verification.
39
AI ResearchAlibaba released Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model, and says open weights are coming — its most capable model yet.
40
AI ResearchWith scaling gains harder to come by, top labs are betting that continual learning and memory architectures — not bigger models — are the path to the next AI leap.
41
AI ResearchDeepSeek's MIT-licensed V4 Flash 0731 lands among the top open-weights models and outscores its larger V4-Pro preview on agent benchmarks through post-training alone.
42
AI ResearchGerman lab Black Forest Labs unveiled FLUX 3, a unified frontier model spanning image, video, audio and action — with a robotics variant already on Audi lines.
43
AI ResearchGoogle rebuilt Gemini 3.5 Pro from the ground up to close reasoning gaps and confirmed Gemini 4 is in pre-training — its most ambitious run yet.
44
AI ResearchNVIDIA and Korea's SK Group unveiled a $500B+ partnership to co-develop next-gen HBM memory and build a 2-gigawatt AI factory powered by Vera Rubin systems.
45
AI ResearchMoonshot AI published open weights for Kimi K3, a 2.8-trillion-parameter mixture-of-experts model that is the largest open-weight system released to date.
46
AI ResearchReuters reports DeepSeek is developing a custom inference chip to reduce its dependence on Nvidia and Huawei, joining a global rush to in-house silicon.
47
AI ResearchAnthropic's Claude Opus 5 posts standout agentic-coding and fluid-reasoning scores at Opus 4.8 pricing, intensifying 2026's frontier-model competition on capability per dollar.
48
AI ResearchMistral is putting a new open-weight model into early access with research, government and industry partners, aiming to close the gap with closed frontier labs.
49
AI ResearchA wave of July frontier launches from OpenAI, xAI and Meta has pushed inference costs to record lows, reshaping how the industry competes on AI.
50
AI ResearchMeta releases Muse Spark 1.1 with a 1M-token context window and computer use, and opens its Model API to developers at roughly a quarter of rival prices.
51
AI ResearchThinking Machines released Inkling, a 975B-parameter Apache-2.0 open-weights model built for customization rather than topping frontier leaderboards.
52
AI ResearchGoogle integrated Gemini Omni and personal AI avatars into Google Vids, enabling conversational video editing with consent-based SynthID watermarking guardrails.
53
AI ResearchChina's Moonshot AI released Kimi K3, a 2.8-trillion-parameter model with native vision and million-token context, pledging to publish weights by July 27.
Fresh AI, agentic AI and robotics coverage — bookmark us and follow along.