Today’s frontier report
Research Pool → Multi-Scout → Editor → Edition
3qk7 · AI LOCAL LABS TECHNICAL SCOUTING

AI FRONTIER

Signals worth understanding before they become obvious.

2026-11-19
01
testinferencecost collapse

180B Parameters, 32GB of RAM: Pushing MoE Inference to the Limit with Transformers

A Hugging Face blog post describes a Transformers implementation that streams only router-selected experts from SSD through a bounded cache, letting Qwen3.8-Flash-Next run on a 32GB Mac at about 14 tok/s and land roughly 6% behind Strata on short decode.

  • Qwen3.8-Flash-Next has 48 layers of 512 experts (10 active), 121 GB of experts in FP8, plus a 51 GB n-gram embedding table.
  • Experts quantized to 3 bits take 49 GB on disk with 12 GiB cached on the device, and dense layers in 8 bits take about 5.4 GB (+0.2% perplexity vs BF16).
  • After loading it uses about 18 GiB, and a 6K prompt peaks at 25.4 GiB.
  • Final results: 200-token decode 13.9 tok/s vs Strata 14.8, 4K prompt prefill 228 tok/s vs 220, decode after a 6K document 12.4 tok/s vs 13.7.

For AI builders, this shows that streaming only routed experts through a bounded slot pool that grouped_mm runs on as is can approach a dedicated engine on consumer memory, while the remaining ~10 ms per token of host round trips is the bot...

Hugging Face Blog
02
testagentscapability unlock

DeltaHarness: Self-Evolving Harnesses for Resource-Adaptive Tool Execution

The paper introduces DeltaHarness, an execution-guided policy-search method that adapts agent harnesses from cold-start measurements; its QHI loop uses observed responses to update runtime dispatch, revise physical plans, and select experiments for later proposals, with an independent confirmation panel controlling pol...

  • With fixed dispatch, a revised materialization policy reduces complete-delivery latency by 30.0% on one video-editing task across three fresh pairs.
  • On held-out tasks, the evolved policies reduce latency by 8.2% relative to the best fixed plan in one seed and by 4.6% relative to the initial proposal in another.
  • The paper reports that the same task and resource description can conceal runtime conditions that reverse the latency ordering of valid execution plans.

For builders of tool-using agents, this suggests the harness itself can be optimized through policy search grounded in measured latency in the target setting rather than relying on static plans.

arXiv cs.LG
03
watchresearchnew architecture

Language Models Can Rewrite Their Own Weights

An arXiv paper reports that a 27B model, given its weights, graded answers and the ability to run itself, diagnoses its failures and edits its own weights with no gradients or training, gaining 6.6 points on 490 unseen problems.

  • The paper reports the edit made the model 6.6 points more accurate on 490 problems it never saw, in a test fixed in advance.
  • The paper reports that over its first eight such sessions, each edit replaced the last and accuracy only rose.
  • The paper reports that repairing a damaged copy, it located the corrupted matrix from text math problems alone and rewrote it, lifting that copy on six unseen image benchmarks from 7% to 70%, 83% of the undamaged model.
  • The paper reports that across four sizes, the 2B model made no edit, the 4B and 9B models broke themselves, and only the 27B model worked; it needs an open model and hours on one GPU.

For builders, this suggests weight-level self-improvement could become a post-deployment improvement path without training infrastructure, though the reported effectiveness so far holds only at 27B scale.

arXiv cs.AI
04
testcreativecapability unlock

Genie 4 is now available

Google DeepMind is rolling out a new Project Genie experience to eligible Google AI Ultra subscribers, giving first-time access to Genie 4 with real-time natural-language environment editing and native audio.

  • Genie 4 lets you prompt mid-session to change weather, drop in a character, or redirect the action in real time.
  • Genie 4 pairs visuals with real-time native audio that adapts as the scene unfolds, from ambient soundscapes to character dialogue.
  • Genie 3 extended sessions to up to 3 minutes per generation, up from 60 seconds, and supports world chaining.
  • Genie 3 adds 360-degree panorama image grounding and real-time in-world promptable events.

For builders, world models are shifting from one-shot generation toward promptable, audio-enabled, shareable interactive environments, while camera latency, intricate physical interactions, and long-term scene permanence remain open challen...

Google DeepMind
05
testmodelsopen source unlock

Rumble-2-Flash: an open 0.5B speech model for 40+ languages

Reson8 open sourced Rumble-2-Flash under the Apache 2.0 license, a 0.5B-parameter speech recognition model for 40+ languages that uses the Qwen3-ASR architecture with native support in Transformers and vLLM.

  • 0.5B-parameter model released under the Apache 2.0 license
  • On the Open ASR Leaderboard multilingual track, it averages a 6.96% error rate across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English
  • On the Open ASR Leaderboard English track it averages 6.6% WER
  • On AppTek's call-centre benchmark (302 hours of helpline calls, 14 English accents) it reaches 7.86% WER, on par with NVIDIA's Nemotron 3.5 Transcribe (30B) at 7.78%

Builders can deploy a 0.5B ASR model under Apache 2.0 with reported multilingual and call-centre accuracy on par with a 30B model, with native Transformers and vLLM support.

Hugging Face Blog
06
watchresearchincremental

Auditors Struggle to Catch Sabotage In End-to-End Transcripts of AI Safety Research

The paper runs a red-team/blue-team game in which a human researcher performs AI safety experiments from scratch as a proxy saboteur while auditors try to tell which experiments were sabotaged and how, given the full research transcript and the ability to run experiments.

  • Human auditors distinguishing sabotaged from benign research achieved an AUROC of 0.59, not significantly better than chance.
  • Human auditors fully identified the specific sabotage in only 17% of cases.
  • The best frontier models, despite not being allowed to run experiments, outperformed human auditors (Fable 5.1: AUROC 0.73, 32% full identification) within noise, and weaker models performed comparably.
  • When the saboteur chose which experiments were attacked rather than being required to sabotage every question given, sabotage identification dropped further.

For teams automating AI R&D, post-hoc auditing of a full research transcript is insufficient to catch expert-level agent sabotage, so audit designs must actively verify experiments.

arXiv cs.AI
07
testdevelopercapability unlock

Inside the Codex sandbox

See how Codex runs commands on your machine while enforcing the filesystem and network limits you control across macOS, Linux, and Windows.

  • Codex runs commands on your machine.
  • Codex enforces filesystem and network limits that you control.
  • The sandbox covers macOS, Linux, and Windows.

Builders can run Codex commands across three major desktop platforms while constraining filesystem and network access to limits they set themselves.

OpenAI News