AI Radar

Top AI stories this week

AI news from the last seven days. Stories that matter more, and that more outlets cover, rank higher. We update this list every hour.

Importance 2 of 5arXiv cs.AI

Study tests whether framing pushes an LLM agent toward market collusion

Researchers placed Gemma 3 12B in a market simulation based on the London silver-fixing case and ran 260 simulations asking it to collude. They varied gain, loss, blame and credit framing plus management pressure to see how it affects manipulation.

Why it matters: Firms deploying agents in finance face risk that persuasive framing can steer agents toward misconduct.

Importance 2 of 5arXiv cs.AI

Audit tests hallucination detectors on small quantized on-device LLMs

Researchers audited four hallucination detectors on small 4-bit models such as Phi-3-mini, Llama-3.2-3B, Qwen2.5-3B and Gemma-2-2B. The study checks whether rankings from large research models hold for models actually shipped on devices.

Why it matters: Teams deploying small on-device models cannot assume detector results from larger models carry over.

Importance 2 of 5r/LocalLLaMA

Developer shows abliteration can be applied as a small runtime LoRA instead of full models

A post explains that most abliteration techniques are rank-1 ablations that fit in a tiny LoRA applied at runtime in llama.cpp. This avoids downloading separate abliterated models and can be toggled per request.

Why it matters: Refusal removal is cheap and easy to apply to open models, which matters for safety controls on self-hosted deployments.

Importance 2 of 5NVIDIA Blog

Nvidia AI Day Seoul highlights Korean firms using its platforms for robotics

More than 2,000 attendees gathered at COEX for Nvidia's AI Day Seoul, focused on physical AI and agents. Nvidia said Doosan Robotics, Hyundai Motor Group, LG Electronics, SK Group and startups are using its platforms to bring general-purpose robotics into real-world use.

Why it matters: It shows Korean industrial groups building robotics on Nvidia's stack, which matters for manufacturers weighing physical AI.

Importance 2 of 5Simon Willison

Simon Willison releases datasette-maplibre 0.1a0, built mostly with one Claude Fable 5.2 prompt

Simon Willison released datasette-maplibre 0.1a0, a Datasette geospatial plugin combining MapLibre GL with the free OpenFreeMap vector map server. He says most of it was built from a single prompt to Claude Fable 5.2.

Why it matters: It is another practical example of frontier coding models producing usable software from a short prompt.

Only want what affects you?
Get a weekly digest filtered to your company. Free.
Filter my radar
Importance 2 of 5OpenAI News

OpenAI launches a science blog on AI-assisted research

OpenAI started a new blog focused on science. It will share findings, practical methods and firsthand accounts from researchers using AI for scientific discovery.

Why it matters: It signals OpenAI's push to position its models as research tools, which matters to R&D-heavy organizations evaluating AI.

Importance 2 of 5r/MachineLearning

Researchers build multiplayer CS:GO-style game rendered by local world models

A team trained a diffusion transformer world model on about 3,958 hours of 10-player CS:GO bot matches and built a multiplayer game where each player's GPU generates their own view. A light server syncs game state; the demo is limited to one small map and pistols.

Why it matters: It is an early sign of world models being used for interactive, shared simulations rather than single-user demos.

Importance 2 of 5r/LocalLLaMA

audio.cpp 0.9.5 adds real-time local avatar voice chat combining three open models

audio.cpp 0.9.5 adds an avatar voice chat app in its WebUI that combines SoulX-FlashHead for talking head video, OWSM v4 for speech recognition and LFM3-Audio for responses and TTS. Video generation still needs an RTX 4090 or better with cuDNN for real-time use.

Why it matters: It shows interactive voice avatars can run fully locally, though hardware demands remain high.

Importance 2 of 5r/MachineLearning

Training-free method edits one weight row so reasoning models stop thinking early

A developer showed that folding a linear probe into a single row of a reasoning model's output layer lets it end its thinking phase once it has reached an answer. The edited model needs no training or extra inference code.

Why it matters: Cutting redundant reasoning tokens could lower inference cost and latency for teams running open reasoning models.

Importance 2 of 5r/LocalLLaMA

User benchmarks 1Cat-vLLM expert caching against Strata on Tesla V100 rig

A user compared inference speeds of 1Cat-vLLM, a Tesla V100 focused fork of vLLM that recently added expert caching, with Strata. The test used a Qwen3.8-Flash-Next IQ3_XXS quant on four 16 GB V100 cards with an Epyc CPU and DDR4 memory.

Why it matters: Expert caching on older GPUs could make large mixture-of-experts models cheaper to run locally on used hardware.

Importance 2 of 5r/LocalLLaMA

VillanovaAI releases Villanova2 open model family for five European languages

VillanovaAI released the Villanova2 family for English, Italian, German, Spanish and French under Apache 2.0. It includes a 15.5B-total, 3.9B-active mixture-of-experts model and a dense 2.4B model, each in base, instruct and vision-language versions, with GGUFs available.

Why it matters: European-built open models give EU companies more options for multilingual work under a permissive license.

Importance 2 of 5r/LocalLLaMA

Open-source tool tunes cudaMemcpy to speed MoE CPU offload prefill by about 35%

A developer released a tool that tunes CUDA memory copies used when offloading mixture-of-experts layers to the CPU. On an RTX 3090 with Qwen3.6 35B-A3B in llama.cpp, prefill rose from 345 to 466 tokens per second.

Why it matters: Cheap software tuning can noticeably improve local inference on consumer GPUs for models too large for VRAM.

Importance 2 of 5r/LocalLLaMA

Basalt 0.4.0 inference engine adds Qwen 3.8 27B, reaching up to 776 tok/s on an RTX 5090

The Basalt local inference engine version 0.4.0 adds its first dense model, Qwen 3.8 27B, with up to 776 tokens per second on structured output on an RTX 5090. Recent updates also added multi-model and router support, DeepSeek V4.1 Flash, video support and files in chats.

Why it matters: Faster local inference makes running capable open models on a single workstation more practical.

Importance 2 of 5r/MachineLearning

Marimba released as open lightweight speech decision model built on Qwen3-ASR-0.6B

A developer released Marimba, a fast, lightweight model that returns typed decisions about speech such as yes/no answers, choices or scores. It runs on a frozen Qwen3-ASR-0.6B, using zero-shot decoder scores for word and language questions and small readouts for voice activity, turn endings, speaker changes and speaker count. Weights are open under Apache-2.0 and a paper is due on arXiv.

Why it matters: Teams building voice agents could use small open components like this for turn detection and speaker handling without large models.

Importance 2 of 5r/LocalLLaMA

Axiomic Labs releases X3Guard-150M, a small prompt guard model for edge devices

Axiomic Labs released X3Guard-150M, a prompt guard model that flags whether a prompt is harmful and whether it is adversarial in a single pass. It is built on the company's GPT-X3 base model and runs in the browser with response times in the tens of milliseconds on iPhone.

Why it matters: Small guard models let teams screen prompts cheaply on device before they reach larger models.