Siri AI developer, Gemini Notebook API, provable inference | TLDR AI. TLDR AI 11/04
Nexus by Fireworks: #1 on the Vals AI Router Index (Sponsor)
Nexus picks the best open or closed model for each turn and learns what your team prefers. It works inside Claude Code, Codex, Cursor and the other tools you already use.
At max effort, Nexus Auto scored 89.3% vs. 86.8% for Claude Opus 5.5 across 1,222 tasks, at 17% lower cost. It's also #1 on the independent Vals AI Router Index, the first router to pass 60%.
New this week: Nexus Desktop for macOS turns it on with one switch. Pro starts at $8/month, with zero data retention on every plan.
Siri AI is now open to all developers (4 minute read)
Developers can now integrate Siri AI into their apps with the iOS 27.3 beta. This update allows apps to leverage Siri's advanced features, like natural language processing and personal context understanding, enhancing user interaction. Siri AI is currently unavailable in the EU and for users under 13. It supports multiple languages, with more to come.
Baseten Introduces Grid, Sandboxes, and Serverless Loops (5 minute read)
Baseten announced Grid for model routing and governance, resumable VM-isolated Sandboxes for agent execution, and general availability of Loops for serverless supervised fine-tuning and reinforcement learning. The products connect training, inference, and execution, with customers retaining ownership of trained weights.
Fireworks Launches Faster AI Inference and Enterprise Tools (5 minute read)
Fireworks AI unveiled new products at Forge, including Ultra-fast inference reaching up to 1,500 tokens per second, lower-cost background processing, and Nexus for accessing multiple AI models within existing tools. It also introduced capabilities for model evaluation, fine-tuning deployment, and enterprise-controlled infrastructure.
How Much Does Reasoning Effort Affect Score and Cost? (8 minute read)
Lower reasoning effort reduces cost much more than it lowers the quality of results. Tests show that models on low reasoning can match models on high reasoning at lower cost. The impact of reasoning effort depends on the task.
Learnings from Auditing Public Evaluations (8 minute read)
The article outlines issues faced during public evaluations of LLMs and details improvements made to benchmarks such as AUTOMATIONBENCH and OFFICEQA PRO V2. They replaced ambiguous prose descriptions with JSON schemas for better performance evaluations and introduced new grading systems to improve accuracy. A multi-layered monitoring system is suggested to prevent manipulation and ensure accurate measurement of model capabilities.
Training a Text-to-Image Model for $21K (13 minute read)
Linium is a two-person research lab that is trying to train its own generative video models entirely from scratch, aiming to create tools to democratize access to animation. Pyramid-JiT (P-JiT) is a pixel-space architecture that achieves aggressive token reduction through a noise schedule that helps pixel-space models better learn fine-grained detail. This study looks at how models trained on the architecture perform when trained for longer as a pressure test along the way towards larger models and video. Examples of images generated by the resulting model are available.
How to Balance AI Value and Risk: Takeaways from the Gartner IT Symposium/Xpo™ 2026 (Sponsor)
AI value is growing, but risk is compounding at a much higher rate. Explore six takeaways from the Gartner IT Symposium/Xpo™ 2026 conference that can help CIOs capture AI value and manage risks through momentum, drag, governance and people. Download the insights.
Gemini Notebook API (1 minute read)
The Gemini Notebook API is a REST interface for programmatically creating and managing notebooks and sources in Gemini Notebook. It can turn web pages, YouTube videos, plain text, and Markdown into persistent notebooks. The API and MCP server are in an early access Public Preview. Only personal Google Accounts are supported during the preview - Google Workspace accounts are not yet supported.
Progress on provable inference for large language models (30 minute read)
It is almost impossible to verify which model produced a specific output. Cryptographic proofs of inference allow model providers to prove that preach suppose was produced which a specific model. These proofs could enable new applications and contribute to a toolkit for verifying compliance with AI pacing requirements. This post details the progress made on a powerful type of cryptographic proof uniquely fit for provable inference called zKSNARK.
Accelerating engineering for the physical world (6 minute read)
OpenAI is leveraging its AI models to enhance mechanical and electrical engineering processes, from design to simulation and production, speeding up internal workflows and improving industry practices. Engineers utilize models like GPT‑6 Astra for tasks such as device design, CAD model creation, and performance simulations, while collaborating with companies like Mazda to explore applications in vehicle aerodynamics.
OpenAI: ChatGPT Search Is About 20% of Google's, 250 Million People Speak With ChatGPT, and Chat Is Merging With Work (16 minute read)
ChatGPT's search volume has reached 20% of Google's, with ChatGPT rapidly gaining 200 million users in the past two months, leading to 1.2 billion weekly active users. The voice mode of ChatGPT now engages 250 million weekly active users, emphasizing new interaction methods through GPT Live updates that enable natural bidirectional conversations. Plans to merge Chat and Work modes aim to simplify user interaction, while Dots highlights innovations in AI task management despite high compute costs.
DeepSeek Plans to Train Its Next-Generation Model on Huawei Ascend Chips (3 minute read)
DeepSeek plans to use Huawei's Ascend AI chips to pretrain DeepSeek V5. Ascend chips have previously been used primarily for inference rather than pretraining. This will be an important test of whether Chinese AI accelerators can support the full pre-training of a frontier large language model. A successful run could reduce Chinese AI developers' dependence on Nvidia hardware and strengthen Huawei's position as an alternative for advanced AI computing.
Your Agents Need Memory. Free Dev Conference, Nov 12 (Sponsor)
Discover agents' durable memory, traceable decisions, and context in production across 100+ technical sessions at Neo4j NODES 2026. Register free.
The 'Astra can read and write weights to improve Qwen' post is misleading (4 minute read)
The 'Astra can read and write weights' claims are misleading.
ChatGPT is about to start making GIFs (1 minute read)
OpenAI just merged animated GIF output into its image generation, with the animations being generated by OpenAI's Images API, which means it is coming to ChatGPT.
Agent State Machines: Building Deterministic Scaffolding for Agents (9 minute read)
This post explains how state machines enforce agent permissions, approvals, retries, and completion criteria outside model prompts.
How much do tokens really cost? (6 minute read)
AI tokens' pricing appears sustainable, as competition among providers is driving down costs.