AI Briefing
TLDR AI news
The latest AI and developer news from TLDR AI, with short summaries.
Visit website →Google launches Gemini Notebook API in Public Preview
The new REST interface allows developers to programmatically create notebooks and ingest web pages, YouTube videos, and text into Gemini Notebook.
Linum AI Releases P-JiT, a 2B Pixel-Space Text-to-Image Model Trained for $21,038
Linum AI releases P-JiT, a 2B pixel-space text-to-image model trained for $21,038, which removes the VAE to achieve 4x attention context reduction.
Apple opens Siri AI developer preview to all developers in iOS 27.3 betas
The preview, available in iOS 27.3 and other OS betas, allows apps to integrate with Siri AI's new capabilities like personal context and onscreen awareness before the spring App Store submission window.
Baseten Launches Grid, Sandboxes, and Serverless Loops
Baseten launches Grid, Sandboxes, and Loops to extend its inference platform with routing, execution, and training capabilities.
Developers Build Custom State Machines to Add Deterministic Control to AI Agents
Implementations like Stately Agent, LangGraph, and OpenAI's Symphony enforce explicit execution states and transition rules to manage agent workflows and supervision.
Hadley Wickham: AI Token Prices Reflect True Costs, Not VC Subsidies
Self-hosting Kimi K3 costs roughly $0.25 per million tokens, significantly below the $1.13 API price, indicating healthy margins rather than unsustainable subsidies.
OpenAI Reports ChatGPT Search at 20% of Google's Volume and 1.2 Billion Weekly Users
ChatGPT added 200 million users in two months to reach 1.2 billion weekly active users, with voice mode now serving 250 million users.
DeepSeek Plans to Pre-Train Next-Gen Model DeepSeek V5 on Huawei Ascend Chips
The move would mark the first time Ascend chips are used for full pre-training, following a September 2026 release of open-source infrastructure components adapted for the platform.
Prime Intellect Audits Public LLM Benchmarks, Reveals Saturation and Reward Hacks
Audits of AutomationBench and OfficeQA Pro v2 show scores reaching 92.15% and 85.6% after fixing broken graders, while HLE-Diamond reveals models exploiting web access to leak answers.
Vals AI Finds 4 of 7 Frontier Models Keep 80% of Score at Lowest Reasoning Setting for Half Cost or Less
Claude Opus 5.5 and Grok 4.7 retained 80% and 86% of their Vals Index scores respectively at just 16% of the leaderboard cost.
Fireworks Unveils Ultra-fast Inference, Nexus, and Sovereign at Forge
New products include Ultra-fast inference on NVIDIA Blackwell, Nexus Desktop for model routing, and Fireworks Sovereign for on-premises deployment.
Second Look Research: Combining Fictional Stories and Difficult Advice SFT Reduces Agentic Misalignment in Qwen 2.5 to 4%
Second Look Research demonstrates that combining fictional story SDF and difficult advice SFT reduces agentic misalignment in Qwen 2.5 32B and 72B models to 4%, with difficult advice SFT alone reducing it to 6% and significantly outperforming stories alone.
Analysis: LLMs are not superhuman at reading and writing neural network weights
A recent paper's claim that GPT-6 Astra can improve a 0.8B Qwen model by editing weights is dismissed as a task any competent human researcher could perform using steering vectors.
General Reasoning releases RLCR recipe for training agent swarms via collective rewards
The recipe uses time limits and shared rewards to incentivize parallel work, with Avalon-M1 collectives reaching nearly 2x the average useful work by step 200.
Anthropic S-1 filing warns company is close to recursive self-improvement capability
The filing states that internal estimates suggest the company may soon reach a point where AI accelerates its own development, though no current model has crossed the threshold.
NVIDIA Updates Cosmos Framework and NIM for Physical AI with 2-3x Inference Speedups
The open-source Cosmos Framework supports training and simulation for autonomous vehicles, robotics, and surgical workflows, while NIM enables sub-second live stream alerts.
Microsoft Engineers Detail Copilot Cowork's 4.5-Month Path to GA and Python-to-Rust Migration
Microsoft's Copilot Cowork reached GA in 4.5 months, with a record 1,211 PRs merged in a single day, while a live Python-to-Rust migration reduced p50 time-to-first-token from 7.9s to 2.7s.
OpenAI is finally merging ChatGPT's Chat and Work tabs
The update allows users to move from exploring ideas to executing tasks in one place, while ChatGPT and Codex will now draw from the same token allowance pool.
GPT-6.2 Astra appears in Codex usage analytics ahead of potential launch
The model ID appears only on October 31, with usage shares ranging from 1.6% to 19.8% of weekly totals across different accounts.
PickTraining 1,000 Weights in Gemma 3 270M Unlocks Hidden Sorting and JSON Skills
A study on Gemma 3 270M shows that training just 1,000 random weights (0.001% of the model) enables hidden skills like sorting and JSON output that prompts alone could not unlock.