The November 2 account links Apache-licensed code and a technical report. The code’s first public commit dates to October 30, and its repository describes an unmaintained research prototype.
Its test uses Llama 4 Maverick’s 400 billion parameters stored at four-bit precision. Weights are the numbers learned during training, streamed here from server memory to one H100 graphics card. The full-model example refuses to start below 330 GB of available server memory.
At 24 requests per batch, the authors report about 110 processed prompt tokens per second. Tokens are the units of information a model processes and generates. Each request contains 1,024 tokens and certifies one next-token choice.
A batch takes about 223 seconds, and each proof occupies 18.78 MB. Checking a proof takes 17 seconds using 24 CPU threads.
Proving costs roughly 70–95 times the GPU time of ordinary prompt processing in the authors’ comparison. The baseline uses a different numerical format on eight H100s.
The verification guide requires a trusted model fingerprint and configuration files obtained separately from the server being checked. Exact checks concern an integer model’s highest-scoring tokens. A looser policy checks only specified score limits.
Caveat: The formal proofs cover a Lean specification, while the fast executable code is tested against it and has had no independent security audit.
(Techmeme · research account, complete technical report and repository limits)
Today’s preview starts in the iOS 27.3, iPadOS 27.3, macOS 27.3 and visionOS 27.3 betas. It uses App Intents, Apple’s framework for exposing app actions.
Developers can provide custom actions and skills that combine several actions. Siri can also surface any entity in Spotlight’s meaning-based search index.
Apple says customer availability starts in a few months, with App Store submissions opening in spring. It will provide further terms and usage policies later.
Some server-based Apple Intelligence features have daily limits. Apple says expanded access will carry a fee in the future, without announcing prices here.
Caveat: This is an early developer preview requiring a supported device and language, and Siri AI remains unavailable in the EU on iOS and iPadOS.
(Techmeme · Apple developer announcement)
Simon Willison’s test gives Qwen 3.8 27B the original prompts containing small game-playing networks’ weights, the numbers learned during training. The prompts prohibit tools and code execution.
Qwen identified Pac-Man in 46 minutes 12 seconds, using 43,506 reasoning tokens. Tokens are the units of information a model processes and generates.
It identified Pong in 1 hour 4 minutes 54 seconds and Breakout in 52 minutes 25 seconds. It misidentified Freeway as a Pac-Man-like game.
The downloadable model file is about 17 GB. Willison used LM Studio and increased its prompt capacity because the default 8,192-token setting failed.
Caveat: Four small game networks and these individual runs provide no general measure of how reliably models can interpret unfamiliar networks.
(HN 359 points · 119 comments · configured X account · published prompts and transcripts)
LlamaIndex reports raising OCRia-1’s text-generation speed from roughly 285 to 2,800 tokens per second on one B200 graphics card.
It combines lower-precision weights with speculative decoding, where a smaller model drafts several tokens for the main model to check together. Sensitive image-processing layers retain higher precision.
Reusing shared preparation reduced roughly 290 small GPU launches per step to one for the model’s recurrent-layer metadata. Other changes reduce memory transfers and accelerate verification of drafted tokens.
The company plans to contribute generally useful changes to vLLM, open-source model-serving software. The article describes OCRia-1 as available across LlamaParse tiers.
Caveat: These are vendor measurements of token generation on a particular workload, so the near-tenfold gain is not a measured tenfold reduction in whole-document processing time.
(Configured X account · full engineering account)
An agent is a model running a multi-step task with tools. The package documents file analysis, transcription, media generation and delegated research for assistants with shell access.
The client can preview requests offline with --dry-run, return structured output and write generated media to disk. Its documentation distinguishes it from Google’s interactive Gemini coding assistant.
The skill warns that GEMINI_API_KEY overrides credentials saved elsewhere. Some run commands exit successfully while reporting an unfinished or failed state, which automation must inspect separately.
Caveat: This beta client calls Google’s hosted API, and its skill documentation says ordinary conversation calls store interaction state on the server by default.
(Configured X account · released skill and client documentation)
Google’s school-search restriction report → UNVERIFIABLE: direct extraction hit an anti-bot block, and the text-proxy fallback required authentication. The policy’s scope remains unchecked beyond collected headlines.
Codex training-setting reports → UNVERIFIABLE: the opened thread contains user reports, but no checked reproduction or provider explanation establishing why the setting changed.
Postman’s automatic plugin installation report → STALE: the detailed evidence dates to October 24–25. Today’s discussion adds no checked new product behavior, maintainer remedy or confirmed data disclosure.