2026

10,781 articles · 136 days · 10 categories · avg 79 per day

Google's StreamVibe video agent beats other models by 10.7% on response timing and quality
LLM·The Decoder

Google's StreamVibe video agent beats other models by 10.7% on response timing and quality

Nov 1, 2026·3 min
New test catches hidden AI agent collusion by measuring how a partner reacts to a paraphrase
Research

New test catches hidden AI agent collusion by measuring how a partner reacts to a paraphrase

Nov 1, 2026·3 min
OpenAI Rewrites GPT Behavior Rules: Five Values and a Chain of Command
Business

OpenAI Rewrites GPT Behavior Rules: Five Values and a Chain of Command

Nov 1, 2026·3 min
Voice bots built on LLMs turn to Jev, a fast classifier, for decisions in 200–500 ms
Business

Voice bots built on LLMs turn to Jev, a fast classifier, for decisions in 200–500 ms

Nov 1, 2026·4 min
MoneyGate safety layer blocked all 42 fraudulent requests to local AI models
LLM·Habr AI

MoneyGate safety layer blocked all 42 fraudulent requests to local AI models

Llama 3.1 8B, Qwen 2.5 7B and DeepSeek-R1 8B on a gaming laptop rarely spotted the scam and wrote to fraudsters, "I'll transfer it today." MoneyGate stopped all 42…

Nov 1, 2026·3 min
Hugging Face's BERT Tokenizer Gets a 50x Speedup by Skipping Needless Work on ASCII
LLM·Habr AI

Hugging Face's BERT Tokenizer Gets a 50x Speedup by Skipping Needless Work on ASCII

The tokenizers team found that BERT spends 88% of its time normalizing text, and taught it to skip English characters in 16-byte batches. The result: a roughly 50x…

Nov 1, 2026·3 min
Google admits AI Mode cuts search revenue, shares fall 15%
Business·Latent Space (swyx)

Google admits AI Mode cuts search revenue, shares fall 15%

Google's chief business officer said surging AI Mode queries hurt revenue because they monetize less. Market cap fell by $500 billion, the worst day since the 2004…

Nov 1, 2026·3 min
Illinois compiled 12 151 open problems and tested whether AI can solve them
Research·arXiv cs.AI

Illinois compiled 12 151 open problems and tested whether AI can solve them

Researchers built a database of open scientific problems and a 200-problem math benchmark. Offline, AI agents solved at most 7 of 40 already-solved problems.

Nov 1, 2026·3 min
Top themes
01Layer Duplication During Training Lifts GSM8K from 12% to 19% on a 3B ModelNov 1, 2026·3 min02ThinkJev Makes AI Think Only About Hard Questions: 28% Fewer TokensNov 1, 2026·3 min03AI engineer writes a single reward function for agent teams: win rate climbs from 51% to 90%Oct 31, 2026·3 min04AI cartoon series: a 24-second episode built ten ways, the best cost 162 creditsOct 31, 2026·4 min05SkillMem stores an AI agent skill in 9.6 KB and speeds up a team 4.2xOct 31, 2026·3 min
Author puts his script before an AI jury: the verdict cost 850 thousand tokens
LLM·TG: @tekhnocafe (Технокафе / Вайбкодинг / AI-агенты)

Author puts his script before an AI jury: the verdict cost 850 thousand tokens

Oct 31, 2026·3 min
Worldwide Express hands disputed invoices to AI agents and cuts unresolved disputes by 75%
Business·AWS Machine Learning Blog

Worldwide Express hands disputed invoices to AI agents and cuts unresolved disputes by 75%

Oct 31, 2026·3 min
Open-source LLM gateway KTMS 1.0 plugs your own GPUs into the cloud and survives outages
Robotics·Habr AI

Open-source LLM gateway KTMS 1.0 plugs your own GPUs into the cloud and survives outages

Oct 31, 2026·4 min
Anthropic: Fold Subagents Back Into a Single Loop — One Customer's Tasks Ran Twice as Fast
LLM·TG: @tekhnocafe (Технокафе / Вайбкодинг / AI-агенты)

Anthropic: Fold Subagents Back Into a Single Loop — One Customer's Tasks Ran Twice as Fast

Oct 31, 2026·3 min
LRMoE trims the MoE backward pass and speeds up training by 21–45%
Research·arXiv cs.AI

LRMoE trims the MoE backward pass and speeds up training by 21–45%

A new MoE training system drops the least useful token–expert pairs in the backward pass. On Qwen3-30B-A3B and GLM4.5-106B-A12B, it is 21.2–44.8% faster than the…

Oct 31, 2026·3 min
The ai-design-check utility cut 266 design issues on a dashboard to zero using 70 rules
Research·X: @sinanyorulmazz

The ai-design-check utility cut 266 design issues on a dashboard to zero using 70 rules

Enmera Software released an open-source checker for AI-agent-built interfaces: 70 rules, no neural network needed, and a skill that fixes the screen and re-verifies…

Oct 31, 2026·2 min
Ontology as a checker for an AI agent: on UniProt, its error flags are only slightly more accurate than random
Research·arXiv cs.AI

Ontology as a checker for an AI agent: on UniProt, its error flags are only slightly more accurate than random

Sano Centre researchers tested an ontology, a formal description of a knowledge graph, as a judge of AI agent answers. On UniProt, the expert ontology flags more correct answers than the others.

Oct 31, 2026·3 min
10 million people have activated Meta's Muse agent, but Hilton and eBay barely see any orders
Autonomous·Wired

10 million people have activated Meta's Muse agent, but Hilton and eBay barely see any orders

Oct 31, 2026·3 min
DeepSeek-V4.1-Flash cuts HBM KV cache to 890 bytes per token, down from 3514
LLM·TG: @gonzo_ML (gonzo-обзоры ML статей)

DeepSeek-V4.1-Flash cuts HBM KV cache to 890 bytes per token, down from 3514

Oct 31, 2026·2 min
OpenAI Seeks $20 Billion at a $1.25 Trillion Valuation Instead of $30 Billion at $1.4 Trillion
Hardware·TG: @addmeto (addmeto (Григорий Бакунов))

OpenAI Seeks $20 Billion at a $1.25 Trillion Valuation Instead of $30 Billion at $1.4 Trillion

Oct 31, 2026·3 min
FAIRouter for C# Is Now on NuGet and Picks a Model From 54 Built-In Options
LLM·Habr AI

FAIRouter for C# Is Now on NuGet and Picks a Model From 54 Built-In Options

Oct 31, 2026·3 min
IBM study: asking the client questions first lifts smart-contract transaction fidelity to 72.9%
Research·arXiv cs.AI

IBM study: asking the client questions first lifts smart-contract transaction fidelity to 72.9%

Oct 31, 2026·3 min
Freefield: an open-source ComfyUI studio with 120 apps for images, video and audio
LLM

Freefield: an open-source ComfyUI studio with 120 apps for images, video and audio

Oct 31, 2026·3 min
22B-Parameter LTX-2.5 Runs on a 10 GB RTX 3080, but Takes 5 Minutes per Shot
Hardware

22B-Parameter LTX-2.5 Runs on a 10 GB RTX 3080, but Takes 5 Minutes per Shot

Oct 31, 2026·3 min
Makers of an AI manager for marketplaces deleted the agent and cut 461 lines of code
Autonomous

Makers of an AI manager for marketplaces deleted the agent and cut 461 lines of code

Oct 31, 2026·3 min
AI Agents Revise IPO Prospectuses in Response to SEC Comments: Learning Adds 3.7 Percentage Points
Business·arXiv cs.AI

AI Agents Revise IPO Prospectuses in Response to SEC Comments: Learning Adds 3.7 Percentage Points

Researchers built FilingLoop from real SEC correspondence. An AI agent that accumulates instructions between tasks meets 27.96% of regulator requirements, versus…

Oct 31, 2026·3 min
Seq@K Metric Across 7 Benchmarks and 16 Models Reshuffles AI Rankings When Feedback Is Available
Autonomous·arXiv cs.AI

Seq@K Metric Across 7 Benchmarks and 16 Models Reshuffles AI Rankings When Feedback Is Available

Researchers at Columbia Business School and UC Santa Barbara propose Seq@K. Across 7 benchmarks and 16 models, rankings with feedback diverge noticeably from the…

Oct 31, 2026·3 min
SciGame Tests Five LLM Agents: Familiar Formulas Get in the Way of Finding Altered Laws
Autonomous·arXiv cs.AI

SciGame Tests Five LLM Agents: Familiar Formulas Get in the Way of Finding Altered Laws

The SciGame benchmark swaps scientific formulas for similar but different ones. Tests on five frontier models show memorized laws hurt agents, and recovering the…

Oct 31, 2026·3 min
Mirendil unveils Fern 0.1: a 10 million-token context window, but a weak agent
Business·TG: @ai_machinelearning_big_data (Machinelearning)

Mirendil unveils Fern 0.1: a 10 million-token context window, but a weak agent

A startup founded by ex-Anthropic researchers unveiled a model that reads 10 million tokens, about 23 MB of code. But it scored just 6.1% on agentic tests, versus…

Oct 31, 2026·3 min
Top themes
01Public AA-Omniscience v1.1 Set Matches the Private One on Net Score with 0.99 CorrelationOct 31, 2026·2 min02Artificial Analysis releases AA-Omniscience v1.1: Net Score drops by about 2 points on averageOct 31, 2026·3 min03Backdoor in Claude extension for Cursor put up to 5,000 developers at risk for nearly 12 hoursOct 31, 2026·2 min04AWS gives coding agents a live site preview straight from an isolated sessionOct 30, 2026·3 min05Mistral opens Large 4 weights: strong at code, but Chinese models score 32 to its 26Oct 30, 2026·4 min
128 AI Agents With No Boss Passed 49% More Tests Than One
LLM·Habr AI

128 AI Agents With No Boss Passed 49% More Tests Than One

Oct 30, 2026·4 min
Claude Sonnet 5 picked one of 19 gene therapies on its own and performed brain surgery on a mouse
Autonomous·The Decoder

Claude Sonnet 5 picked one of 19 gene therapies on its own and performed brain surgery on a mouse

Oct 30, 2026·3 min
Hyperflow H3 charges $9 per minute of image-to-video, while Dreamina Seedance 2.5 asks $34.12
Creative·X: @ArtificialAnlys

Hyperflow H3 charges $9 per minute of image-to-video, while Dreamina Seedance 2.5 asks $34.12

Oct 30, 2026·2 min
Hyperflow H3 Ranks 5th in AA-Video-I2V and 3rd in AA-Video-Editing
Creative·X: @ArtificialAnlys

Hyperflow H3 Ranks 5th in AA-Video-I2V and 3rd in AA-Video-Editing

Oct 30, 2026·2 min
REA runtime lifts robot success on LIBERO-Pro to 95.9% and cuts control steps by 25%
Robotics·Hugging Face Daily Papers

REA runtime lifts robot success on LIBERO-Pro to 95.9% and cuts control steps by 25%

REA splits a robot into a "brain" and fast reflexes. Across 3000 LIBERO-Pro episodes it reached a 95.9% success rate vs. 94.7% for the best single agent, with 25%…

Oct 30, 2026·3 min
T-Bank Fine-Tunes LLMs With Gradient-Free Evolution: 7.8x Faster on 128 GPUs
Research·TG: @cgevent (Метаверсище и ИИще)

T-Bank Fine-Tunes LLMs With Gradient-Free Evolution: 7.8x Faster on 128 GPUs

T-Bank's ESSA method fine-tunes an LLM without gradients. It changes a million times fewer parameters than GRPO, matches it on math, and runs up to 7.8x faster on…

Oct 30, 2026·3 min
Inclusion AI speeds up LLM generation by up to 6.9x with a draft tree
LLM·TG: @ai_machinelearning_big_data (Machinelearning)

Inclusion AI speeds up LLM generation by up to 6.9x with a draft tree

Inclusion AI's UBTree builds a tree of 64 candidate continuations, and the large model verifies it in a single pass. Average speedup on Qwen3-8B: 6.94×.

Oct 30, 2026·3 min
Strata Runs Qwen3.8-Flash-Next on a 12 GB RTX 5070: 94 Tokens per Second at Q2_0
Hardware·Habr AI

Strata Runs Qwen3.8-Flash-Next on a 12 GB RTX 5070: 94 Tokens per Second at Q2_0

Oct 30, 2026·3 min
Agent traffic on Zocdoc is up 30x since July, and ChatGPT faces two health lawsuits
Autonomous·Habr AI

Agent traffic on Zocdoc is up 30x since July, and ChatGPT faces two health lawsuits

Oct 30, 2026·3 min
Credit Saison calculates 954% ROI from a ChatGPT Enterprise pilot with 315 employees
Business·OpenAI Blog

Credit Saison calculates 954% ROI from a ChatGPT Enterprise pilot with 315 employees

Oct 30, 2026·3 min
Mistral Releases Large 4 With Open Weights: 1T Parameters, 52B Active
Research·X: @Z3nlyte

Mistral Releases Large 4 With Open Weights: 1T Parameters, 52B Active

Oct 30, 2026·3 min
MoEngage Opens 400 Private APIs to AI Agents Through AgentCore Gateway, Without Building Its Own MCP Servers
Business·AWS Machine Learning Blog

MoEngage Opens 400 Private APIs to AI Agents Through AgentCore Gateway, Without Building Its Own MCP Servers

Oct 30, 2026·4 min
TELUS voice agent built on ElevenAgents more than halves 30-day cancellations
Autonomous

TELUS voice agent built on ElevenAgents more than halves 30-day cancellations

Oct 30, 2026·3 min
Open-Weight GLM 5.3 Built Firefox Exploits Without Human Help, at About $1,000 per Working Attack
LLM

Open-Weight GLM 5.3 Built Firefox Exploits Without Human Help, at About $1,000 per Working Attack

Oct 30, 2026·3 min
baton utility moves a session between 5 CLI AI agents without losing context
Autonomous

baton utility moves a session between 5 CLI AI agents without losing context

Oct 30, 2026·3 min
Hard Stop halts an AI agent in the Linux kernel with a median latency of 4.8 microseconds
Autonomous·Habr AI

Hard Stop halts an AI agent in the Linux kernel with a median latency of 4.8 microseconds

Hard Stop places a supervisor between an AI agent and the Linux kernel: any action outside the allowlist is blocked, with a median trigger latency of 4.8…

Oct 30, 2026·3 min
Claude Code built a FLUX.3 klein pipeline: 290 doctor images for $3.2 in GPU rental
LLM·Habr AI

Claude Code built a FLUX.3 klein pipeline: 290 doctor images for $3.2 in GPU rental

A paid-traffic specialist for clinics had Claude Code build a pipeline: a website link in, a folder of vetted doctor images out. The result: 290 images for about…

Oct 30, 2026·3 min
Tencent open-sources Hy-Nomas world model: 18.5 FPS on 8×B200 with full history
Creative·TG: @ai_newz (эйай ньюз)

Tencent open-sources Hy-Nomas world model: 18.5 FPS on 8×B200 with full history

Tencent has released the code and weights of its Hy-Nomas world model under Apache 2.0: text and camera control, LoRA fine-tuning, and about 18.5 FPS on 8×B200.

Oct 30, 2026·2 min
Pika Open-Sources H3 Flex Video Model: 15 Seconds With Audio in 6 Steps
LLM·TG: @neurogen_news (Neurogen)

Pika Open-Sources H3 Flex Video Model: 15 Seconds With Audio in 6 Steps

Pika has released open weights for H3 Flex: up to 15 seconds of video with built-in audio in 6 denoising steps instead of 50. It's realistic on a home GPU with 24 GB of memory.

Oct 30, 2026·3 min
Google: AI Agents Using Android CLI Complete UI Tasks 77% Faster
Business·TG: @ai_machinelearning_big_data (Machinelearning)

Google: AI Agents Using Android CLI Complete UI Tasks 77% Faster

Oct 30, 2026·3 min
Microsoft's TokenEconomizer Plugin Cuts Coding Agent Costs by 18.4–30.9%
Autonomous·TG: @notboring_tech (Not Boring Tech)

Microsoft's TokenEconomizer Plugin Cuts Coding Agent Costs by 18.4–30.9%

Oct 30, 2026·2 min
At RGS, local LLMs cut the review of 500 tasks from five working days to two
Business·Habr AI

At RGS, local LLMs cut the review of 500 tasks from five working days to two

Oct 30, 2026·3 min
HOSTKEY wrote a pipe for its bot: Gemma 4 answers in 5–7 seconds, Qwen 3 in 30–40
Autonomous·Habr AI

HOSTKEY wrote a pipe for its bot: Gemma 4 answers in 5–7 seconds, Qwen 3 in 30–40

Oct 30, 2026·4 min