LFM3 launches six specialized models around two billion parameters, sharing a common backbone. They cover text, vision, audio, combined inputs and two decision variants.
The text model card specifies 1.73 billion parameters and a 262,144-token window for input plus output. Tokens are the text units a model processes.
Its four-bit model uses about 1 GB before the additional memory needed for long conversations. Lower precision stores weights more compactly; the card lists llama.cpp, MLX and GPU-serving options.
Liquid trained it for Hermes, Pi and OpenClaw. An agent is a model running a multi-step task with tools. Deferred tools let the surrounding software reveal function definitions after searching a catalog.
In Liquid’s browser-task comparison, it scores 38.42% versus Qwen3.5-9B’s 28.94%. On repository repairs, it scores 43.40% versus 51.20%, showing why one task’s advantage cannot establish a universal winner.
Open-weight means downloadable model, with licence terms attached. Its LFM Open License includes a $10 million annual-revenue threshold for commercial use.
Caveat: These are vendor comparisons, and the model card recommends against autonomous coding and complex multi-step reasoning with this small text model.
(HN 66 points · 24 comments)
Interactive generated worlds add sound, with access still limited#
Genie 4 generates an interactive environment as keyboard actions and text instructions arrive. It adds speech, music and ambient sound alongside the generated scene.
Gemini models help plan events and monitor incoming frames. Google’s examples use Gemini 3.8 Flash to adjust plans as the user intervenes.
Trusted testers have preview access. Project Genie access for Google AI Ultra subscribers and early-access API partners is described as coming soon, with a waitlist available.
Google describes creative prototyping and simulated training environments as applications. Its release also acknowledges drifting objects, unreliable text, physical inconsistencies and difficulties maintaining longer scenes.
Caveat: These demonstrations establish neither a generally available developer service nor a validated substitute for a game engine or physical simulator.
(HN 495 points · 296 comments · Techmeme · configured X announcement)
AI bug searches and generated-input testing find different failures#
Project Zero’s October 27 comparison examines V8, Chrome’s JavaScript engine. Fuzzilli generates test programs, while BigSleep reviews code changes and writes tests that demonstrate crashes.
From April through September, BigSleep produced 45 actionable vulnerability reports and Fuzzilli 37. These counts require associated fixes and exclude duplicate or intended-behavior reports.
Observed production overlap is roughly 25%. Once one system’s discovery is fixed, the other loses its opportunity to rediscover it, limiting that comparison.
A separate experiment gives BigSleep 25 attempts on each of 112 selected historical bugs. It rediscovers roughly 62%; changed prompts and starting points raise the combined result to about 70%.
Bugs that were difficult for Fuzzilli also tended to be difficult for BigSleep. The study supplies concrete examples involving unusual inputs, unsupported language features and interactions between distant compiler stages.
Caveat: This compares two systems on one heavily tested engine, and the authors leave cost per vulnerability unresolved rather than establishing AI’s economic superiority.
(lobste.rs · production records, rediscovery methods and case studies)
Apple plans direct document access for external assistants#
October 27’s announcement adds MCP, the standard for connecting agents to tools. External assistants can read content, edit documents and execute actions within the three apps.
MCP support requires macOS 27 or later. The built-in document assistants separately require Apple Intelligence-capable devices and are initially available in U.S. English.
The built-in assistants combine Apple and OpenAI models. Their advertised tasks include restructuring writing, refining slides, organizing spreadsheet data and suggesting charts.
Apple says the Creator Studio updates arrive in November as a free update for existing subscribers. This is announced availability, distinct from the collected report’s present-tense wording.
Caveat: Apple’s announcement does not document the external tools’ approval boundaries or establish that connected third-party assistants process documents entirely on-device.
(Techmeme’s collected report · first-party announcement and requirement footnotes)
Measure delivered work rather than rewarding model usage#
The memo prioritizes long-term customer and business outcomes. It warns that token consumption, adoption rates and code-change counts can reward spending and complexity when turned into targets.
It still permits these measures with judgment and recommends examining both low-adoption and heavy-use teams. Claims that higher adoption correlates with better delivery appear without supporting measurements in the memo.
Engineers are asked to document tasks where current AI tools provide little benefit. Leaders should be able to describe production tasks where tools helped and where they failed to help.
Caveat: This is an employer’s stated guidance, not an independent productivity study or evidence that its performance reviews consistently follow the memo.
(Techmeme · original reporting checked against the newly public memo)
Bruce Gil’s October 27 report follows a newly public account from Grace Knowles. Her reported stop occurred in September 2025, rather than this week.
In Jay Stanley’s October 26 ACLU account, Knowles says an officer showed her a map containing dated plate photographs. He attributed the stop to suspicious travel patterns.
She says the officer questioned her New Orleans visit because his system showed no sightings there. Missing camera evidence was treated as a reason to doubt her explanation.
The ACLU connects this account with previously reported stops and broader plate-data analysis. Its objection concerns generating suspects from movement histories, beyond looking up vehicles already connected to an investigation.
Caveat: The account leaves the specific software and the officer’s agency independently unverified, so it cannot identify which model or agency made this decision.
(HN 270 points · 199 comments · linked firsthand account and reported analysis)
The tokenizer-cost report → STALE: its underlying CLEAR analysis is dated September 29. The useful language-cost comparison supplies no checked new model evaluation in this report.
The Llama menu-bar demonstration → THIN_GUARD: the recovered original post and actual application site describe a usable local server. Today’s short demonstration supplies no separately verified release or feature change.
Oracle’s expanded AI partnership → THIN_GUARD: new integration and accelerator plans are real announcement subjects, but the inspected report lacks concrete rollout dates and deployment instructions.
Wikipedia’s cleanup report → UNVERIFIABLE: extraction returns navigation and comments rather than the article. Its headline alone cannot support a factual item.
The engineering-knowledge essay → UNVERIFIABLE: direct extraction was blocked and the alternate reader returned 401. No argument or conclusion is inferred from its title.