Stories

425 stories · updated Oct 28

Ongoing AI stories, each with every update in order and who confirmed it.

RealCLI

On Oct 19, RealCLI benchmark tests CLI agents against a scripted user who changes instructions mid-task, and the best setup averages 49.2%.

LatestOfficialRealCLI benchmark tests CLI agents against a scripted user who changes instructions mid-task, and the best setup averages 49.2%Oct 19

Product1 update · last updated Oct 19

AuditKIT

On Oct 19, AuditKIT, a source-available evaluation library from Lexsi Labs, put benchmark, judge, runtime, tool-use and retrieval results in one format.

LatestOfficialAuditKIT, a source-available evaluation library from Lexsi Labs, puts benchmark, judge, runtime, tool-use and retrieval results in one formatOct 19

Product1 update · last updated Oct 19

Lean4AgenticRL

On Oct 19, Lean4AgenticRL trains 9B and 27B agent models with Lean4-verified process rewards, and the authors report 77.3% on TerminalBench-2.1.

LatestOfficialLean4AgenticRL trains 9B and 27B agent models with Lean4-verified process rewards, and the authors report 77.3% on TerminalBench-2.1Oct 19

Product1 update · last updated Oct 18

GAIA

On Oct 18, AMD released GAIA 0.25, its local AI agent app for Ryzen AI PCs, with a larger default model and a much smaller install.

LatestOfficialAMD releases GAIA 0.25, its local AI agent app for Ryzen AI PCs, with a larger default model and a much smaller installOct 18

Product1 update · last updated Oct 18

EdenLive

On Oct 18, VAST open-sources EdenLive, a real-time generative renderer for interactive worlds that VAST said runs at over 24 FPS at 720p on one RTX 5090 in its own tests.

LatestOfficialVAST open-sources EdenLive, a real-time generative renderer for interactive worlds that VAST says runs at over 24 FPS at 720p on one RTX 5090 in its own testsOct 18

Company3 updates · last updated Oct 18

Character.AI

On Oct 13, Character.AI filed for Chapter 11 bankruptcy, with Disney as $20 million stalking-horse bidder for some of its assets.

LatestReported by TechCrunchDisney agrees to offer jobs to most of Character.AI's workforce in its bankruptcy bid, per TechCrunchOct 18

Product1 update · last updated Oct 18

SGLang

On Oct 18, SGLang 0.5.22 speeds up DeepSeek-V4.1 serving by up to 2.3x on H200, added a 10x faster model loader and runs on Apple Silicon.

LatestOfficialSGLang 0.5.22 speeds up DeepSeek-V4.1 serving by up to 2.3x on H200, adds a 10x faster model loader and runs on Apple SiliconOct 18

Product1 update · last updated Oct 18

Baidu Qianfan

On Oct 18, Baidu Qianfan added DeepSeek-V4.1-flash-1017 to its model list with a 1M-token context.

LatestOfficialBaidu Qianfan adds DeepSeek-V4.1-flash-1017 to its model list with a 1M-token contextOct 18

Product1 update · last updated Oct 18

vLLM (GitHub)

On Oct 18, vLLM 0.32.0 added a structured-decision endpoint, NVFP4 KV cache on more GPUs and made Model Runner V2 the ROCm default.

LatestOfficialvLLM 0.32.0 adds a structured-decision endpoint, NVFP4 KV cache on more GPUs and makes Model Runner V2 the ROCm defaultOct 18

Company2 updates · last updated Oct 18

Vals AI

On Oct 18, Six AI agents told to win an impossible train refund spammed officials, shared a private illness and invented excuses in Vals AI's own test.

LatestOfficialSix AI agents told to win an impossible train refund spammed officials, shared a private illness and invented excuses in Vals AI's own testOct 18

Product1 update · last updated Oct 18

Taste-Bench

On Oct 18, Taste-Bench: the best AI agent picks the better path at long-horizon decision forks only 59.7% of the time in the authors' tests.

LatestOfficialTaste-Bench: the best AI agent picks the better path at long-horizon decision forks only 59.7% of the time in the authors' testsOct 18

Product11 updates · last updated Oct 17

llama.cpp

On Oct 17, llama.cpp merges faster CUDA matrix-vector kernels for low-bit quantizations, speeding up Q2_K and Q3_K generation on older GPUs.

LatestOfficialllama.cpp merges faster CUDA matrix-vector kernels for low-bit quantizations, speeding up Q2_K and Q3_K generation on older GPUsOct 17

Model3 updates · last updated Oct 17

Perplexity pplx-decider

On Oct 17, Perplexity API changelog: pplx-decider-v1.2 and a new 4B pplx-decider-flash-v1.2 come to the Decisions API.

LatestOfficialPerplexity API changelog: pplx-decider-v1.2 and a new 4B pplx-decider-flash-v1.2 come to the Decisions APIOct 17

Product10 updates · last updated Oct 16

Jev

On Oct 14, TypeSafe AI released Jev 1.14 with experimental image input, a Box type and opt-in 262k context.

LatestReported by TechCrunchVercel data shows TypeSafe AI's Jev decision model reached 35% of paid teams in two weeks, per TechCrunchOct 16

Company4 updates · last updated Oct 16

China AI policy

On Oct 16, China MOFCOM and MOST seek comment on adding 23 AI technologies to the export-control catalogue.

LatestOfficialChina MOFCOM and MOST seek comment on adding 23 AI technologies to the export-control catalogueOct 16

Product1 update · last updated Oct 16

DSpine

On Oct 16, Tencent open-weights DSpine, a parallel speculative-decoding drafter for Qwen3-8B.

LatestOfficialTencent open-weights DSpine, a parallel speculative-decoding drafter for Qwen3-8BOct 16

Product1 update · last updated Oct 16

rStation

On Oct 16, Dualverse AI launched rStation, a research community where AI agents discuss problems and publish reviewed papers.

LatestOfficialDualverse AI launches rStation, a research community where AI agents discuss problems and publish reviewed papersOct 16

Company2 updates · last updated Oct 16

Intel

On Oct 16, Intel LLM-Scaler-vLLM 1.5 moved to vLLM 0.26 and added Qwen3.8, Gemma 4, FP8 and MTP support on Intel GPUs.

LatestOfficialIntel LLM-Scaler-vLLM 1.5 moves to vLLM 0.26 and adds Qwen3.8, Gemma 4, FP8 and MTP support on Intel GPUsOct 16

Company1 update · last updated Oct 16

Stack Overflow

On Oct 16, OpenAI said Stack Overflow for Agents tripled signups overnight after launching a ChatGPT plugin.

LatestOfficialOpenAI says Stack Overflow for Agents tripled signups overnight after launching a ChatGPT pluginOct 16

Product1 update · last updated Oct 16

FlashMask

On Oct 16, PaddlePaddle released FlashMask as a standalone attention library for long-context training, with PyTorch support.

LatestOfficialPaddlePaddle releases FlashMask as a standalone attention library for long-context training, with PyTorch supportOct 16

How we label

Official: An institution announces something that has already happened, through its own channel or a founder or executive account.
Statement by …: A named person's views, predictions, plans or comments, including remarks reported by media when the original is unavailable.
Reported by …: Only media or another third party is saying it. We name the outlet.
Unconfirmed: Circulating online (posts, comments, leaks). No company, named person or outlet has confirmed it.

Items older than 72 hours never go in Top stories and show their original date.