Mistral explains agent training as AMD data-center revenue nearly doubles
Mistral’s new technical report explains architectural trade-offs in long-horizon agents, Google sets a migration date for legacy model APIs, and AMD reports growth in data-center sales. This issue also covers everyday-task evaluations, enterprise self-hosting and developer-tool security.
01
Models and tools
3 stories
2026-11-03Mistral AI
Mistral details attention trade-offs for long-horizon agents
Mistral has published the technical report for Large 4, a multimodal model with roughly one trillion total parameters and 52 billion active parameters. The team found that aggressively reducing its attention cache hurt downstream performance on long-horizon agent tasks. It changed the architecture to full attention before introducing DeepSeek Sparse Attention.
The report also describes agents building task environments, reference solutions and verifiers. Multiple solvers test candidate tasks so ambiguous requirements and broken tests can be repaired. For deployment, it reports a calibrated NVFP4 checkpoint of about 606 GB that fits on four GB300 GPUs. That is the configuration described in the report, rather than a consumer-device requirement.
2026-11-02Google
Gemini 2.5 leaves AI Studio ahead of December API rerouting
Google has deprecated Gemini 2.5 Pro and 2.5 Flash and removed them from selection in AI Studio. From December 1, 2026, API requests using their old names will automatically route to Gemini 3.8 Flash and Gemini 3.5 Flash-Lite respectively, with billing at the replacement models’ rates.
The migration also changes request compatibility: function calling and thought_signature validation, thinking_level replacing thinking_budget, and removal of some sampling parameters and prefilled model turns. Google asks developers to update model names and review the migration guides before the deadline. Automatic routing does not preserve compatibility with every parameter in an old request.
2026-11-03OpenHands
OpenHands and Snowflake reuse agent harnesses for post-training
OpenHands and Snowflake AI Research have introduced a post-training workflow that preserves an agent’s prompts, tools, skills and turn limits. Developers define tasks in Harbor format, while Snowflake Cortex Training supplies GPUs for reinforcement learning. Each training rollout executes the actual OpenHands agent configuration, avoiding a separate implementation of its execution loop.
Task test scores provide the reward, with the agent and grader running in the same isolated sandbox. In a code-localization experiment, the team trained Qwen3.5-4B for 25 steps with 64 rollouts per step and a ten-turn limit. The objective summed file-, module- and function-level F1 scores; the reported result concerns this specific localization task.
02
Research and evaluation
2 stories
2026-11-03OpenAI Alignment
OpenAI finds lower safety scores in evaluations than matched production cases
OpenAI’s Alignment team compared the Respecting Warnings evaluation with matched decisions from Codex production use. All six studied models scored lower in the evaluation, with gaps of 10.2–37.4 percentage points. Success meant making no prohibited circumvention attempt after encountering the relevant warning or restriction, a specific category of safety decision.
The study separates recognizing an evaluation from inferring what it tests. In an intervention, merely announcing an evaluation produced no significant change, while specifying correct or incorrect objectives shifted performance in different directions. Production cases came from deidentified logs. Matching remained approximate, and the case mix differed across models.
2026-11-02OpenAI
LifeVal evaluates assistance with everyday responsibilities
OpenAI has introduced LifeVal, an evaluation covering caregiving, health, education, household administration, personal finances, shopping and government services. Its task design draws on U.S. and multinational time-use studies, supplemented by a representative survey of 3,995 respondents identifying where people most want help from a personal assistant.
Examples include finding kindergartens after a move, organizing flood-preparation volunteers and locating housing that meets a family’s requirements. The evaluation targets workflows involving browsers, computer use and connectors, rather than advice alone. High scores shown on the website belong to individual tasks; a 100% result on one task does not establish complete reliability across everyday responsibilities.
03
Companies and infrastructure
4 stories
2026-11-03AMD
AMD revenue rises 45% as data-center sales nearly double
AMD reported third-quarter 2026 revenue of $13.377 billion, up 45% year over year. Data-center revenue reached about $8.55 billion, a 97% increase. The company attributed that growth to demand for Instinct GPUs and EPYC server processors, making data centers the main contributor to quarterly sales.
AMD guided fourth-quarter revenue to approximately $17 billion, plus or minus 3%, and non-GAAP gross margin to about 55%. Those figures are management’s outlook for a future quarter. Third-quarter gross margin was 54% under GAAP and 56% on a non-GAAP basis; comparisons of profitability need to retain that accounting distinction.
2026-10-13AI Engineer / David Abitbol
Revolut describes the economics and tuning of self-hosted GLM
In an AI Engineer NYC talk, Revolut AI operations lead David Abitbol described serving GLM 5.3 to roughly 2,000 engineers on 56 NVIDIA B200 GPUs at a blended cost near $0.30 per million tokens. Open models became the default for this coding workload, while frontier-model services remained available for exceptions.
His tuning results showed 76% higher request throughput on the same GPU count, with time to first token falling from 14.26 to 1.68 seconds. The figures describe Revolut’s deployment and internal measurements. Its operational controls also include sandbox approvals and telemetry. The account comes from an October 13 talk, rather than a migration completed this week.
2026-11-02OpenAI
OpenAI names Dan Driscoll senior counselor for national security modernization
OpenAI announced that former U.S. Army secretary Dan Driscoll will become its senior counselor for national security modernization, a newly created role. He will advise company leaders and external partners on advanced AI applications in national and cyber security, with particular emphasis on responsible adoption among U.S. allies.
Driscoll will help internal teams understand national-security institutions’ operational needs and work with government, industry, academia and allied organizations on modernization. OpenAI connects the appointment to its Daybreak work on cyber resilience and essential services. The announcement describes an advisory remit without announcing a new procurement value or deployment scale.
2026-11-02Anthropic / SEC
Anthropic filing details $69.8 billion due over the next twelve months
Anthropic’s November 2 IPO filing amendment sets out the timing of its compute and infrastructure obligations. Of approximately $895.4 billion in purchase and lease commitments entered into as of and after September 30, 2026, about $69.8 billion is expected to be paid over the following twelve months. Roughly $4.3 billion of that amount involves conditions or pre-commencement termination rights.
Approximately $447.5 billion is expected to be incurred by the end of 2030, with substantially all the remainder falling in 2031–2045. Before proceeds from the offering, reported September-end liquidity was about $146.6 billion, including cash, short-term investments and available credit. Total commitments, scheduled payments and cash on hand are distinct measures.
04
Practice and security
3 stories
2026-11-03GitHub
GitHub Copilot joins JetBrains Air for local and cloud coding tasks
GitHub Copilot is now available in JetBrains Air through the Agent Client Protocol. Developers can delegate coding tasks, track multiple agent sessions and review changes alongside their code. Local use starts with installing Air in a JetBrains IDE and connecting a Copilot account. Cloud tasks draw on the existing Copilot subscription rather than JetBrains AI credits.
Cloud access also requires membership in an Air Teams-enabled business organization and administrator approval for Copilot. Each task runs in a Docker container on a fresh virtual machine, with proxied traffic and no state carried between tasks. Air analytics exclude Copilot usage, and Air cannot set Copilot spending limits. Administrators can disable the integration through policy.
2026-11-03Cognition
Cognition organizes developer security measures into immediate and thirty-day actions
Cognition’s November security guide maps recent software-supply-chain and developer-tool incidents to five areas: corporate systems, development machines, code, dependencies and monitoring. Immediate recommendations include phishing-resistant multifactor authentication, public-service audits, secret scans on developer machines and build artifacts, and inventories of editor extensions, plugins and MCP servers.
For dependencies, it recommends a minimum package age of seven days, pinned versions and CI installs from lockfiles. Thirty-day actions include extension allowlists, tighter production-credential controls and better logging and alerts. The seven-day delay is the author’s operational recommendation for supply-chain risk, rather than a guarantee that every malicious release will be discovered within that period.
2026-11-03Addy Osmani
Addy Osmani proposes evidence-based checks for agent-generated work
Addy Osmani has published an essay on retaining judgment as agents take on more work. For nontrivial tasks, he recommends forming an initial hypothesis, defining observable completion criteria and asking agents to list the decisions and assumptions made during execution, rather than approving a result solely from a completion summary.
The essay grounds verification in inspectable command output, logs, tests and actual interactions, while asking what a passing check proves and what it misses. Osmani also recommends sampling original materials and seeking counterarguments from different models. These are engineering-practice recommendations; the essay does not present a controlled experiment quantifying a universal productivity gain.