Mistral has raised €3 billion in a Series D led by Samsung, with funding earmarked for frontier research, training compute, infrastructure and commercialization. NVIDIA RTX Spark laptops are now available, offering up to 128GB of unified memory and 1 petaflop of FP4 AI compute to run local models alongside creative projects. DeepSeek’s Search and Fetch APIs cost $0.75 per 1,000 successful requests, with no charges for failures, empty searches or timeouts. OpenCode Console was breached, exposing users’ own model-service credentials; users who only use the coding agent or desktop app are unaffected.
My take
To me, today's bigger story isn't model performance but the limits of control. Mistral is emphasizing autonomy, while the OpenCode Console breach exposed users' own credentials: owning the account doesn't mean you're the only one holding the keys.
Mistral is expanding its training infrastructure, while RTX Spark brings local AI capabilities to laptops.
Mistral announced a €3 billion Series D led by Samsung Electronics, at a post-money valuation of over €21 billion. The funding will support frontier research, expanded training compute and infrastructure, and faster commercialization and international growth. Following its ASML-led Series C, Mistral has once again secured an industrial technology company as its lead investor. The company says enterprise and government demand is shifting from pursuing the strongest model to balancing performance with control over data, models and compute.
Read original →NVIDIA announced the availability of RTX Spark Windows laptops from six manufacturers, including ASUS and Dell. They integrate a Blackwell RTX GPU and a Grace CPU into a single superchip, offering up to 128GB of unified memory and up to 1 petaflop of FP4 AI compute. Shared memory allows local models and creative projects to run simultaneously. Agents such as Hermes and OpenClaw offer guided installation with no command line required.
Read original →🔍 Analysis: Mistral’s emphasis on control over data, models and compute, and RTX Spark’s 128GB of unified memory, point to the same technology-selection question: which capabilities need to remain in your own hands? They operate at different levels: the €3 billion funding round supports a model developer’s expansion, while local hardware gives users another place to run models. For enterprises, deployment location and control over resources need to be procurement criteria in their own right, alongside model performance. The ability to run a model locally does not, by itself, mean having control over the model itself.
Replit names its preferred partner, DeepSeek adds retrieval APIs, and Sphinx makes code comprehension a prerequisite for pushing changes.
DeepSeek’s Search and Fetch APIs target agents and RAG applications, charging $0.75 per 1,000 successful requests. Search uses an in-house engine and returns structured results, while Fetch converts a single webpage into Markdown. Billing is based on neither tokens nor the number of results returned, and failures, empty searches and timeouts are free. Webpage fetching is subject to the target site’s availability and access policies.
Read original →Replit CEO Amjad Masad clarified that OpenAI has become its preferred partner for powering Agent. He explained that his remarks at TC Disrupt reflected model performance, cost and the relationship between the two companies. Replit remains model-neutral: users can choose Anthropic and other supported models. Preferred-partner status does not mean exclusive use.
Read original →Sphinx is a Git hook for agent-assisted development that pauses pushes to protected branches and asks questions to check whether developers understand the code. The review model does not read the preceding chat; instead, it asks follow-up questions about edge cases based on the code diff. Wrong answers trigger an explanation and a new question, while correct answers allow the push to proceed. It works with tools including Codex, Claude Code and Cursor. Answers are not passed to the coding agent, and it does not read .env files or secrets. Checks can be skipped with sphinx yolo, but every bypass is logged.
Read original →🔍 Analysis: Replit favors OpenAI while retaining other model options, and Sphinx works across Codex, Claude Code and Cursor. This suggests that model partnerships and control over development workflows do not have to be tied together. DeepSeek’s retrieval APIs, priced at $0.75 per 1,000 successful requests, also make the cost of accessing external information separately measurable. For development teams, model performance, retrieval costs and responsibility for code delivery should be evaluated separately. Choosing a preferred model cannot substitute for validating the entire agent workflow.
The scope of the OpenCode Console breach must be assessed separately from the coding agent itself.
OpenCode Console was breached, exposing model-service credentials supplied by users. Console is a hosted inference gateway similar to OpenRouter, not the same product as the OpenCode coding agent. This incident affects Console; users who only use the CLI/TUI coding agent or desktop app are unaffected.
Read original →Atlas turns potential warehouse worker actions into risk scores; Amazon denies monitoring individuals.
Internal Amazon documents show that its in-house AI tool, Atlas, is used to predict the likelihood of warehouse workers in the US, Canada and the UK taking actions such as organizing unions or staging sit-ins. The tool aggregates hundreds of data inputs, including Reddit posts and employee feedback surveys, to assign warehouses risk scores from 100 to 900, with lower scores indicating a greater likelihood of action. Amazon denies monitoring individuals, saying the analysis is conducted only at the warehouse level and does not track individual opinions or union membership. It says the aim is to understand what support employees need.
Read original →An Anthropic researcher uses conceptual representations and implicit planning to explain the limits of describing models as “next-token predictors.”
In an interview with Lenny, Anthropic interpretability researcher Emmanuel Ameisen discussed the internal mechanisms of models such as Claude and their connection to AI safety. He explained that models develop general conceptual representations and can plan multiple tokens ahead without explicitly writing out those plans. In his view, describing a model simply as a “next-token predictor” understates how it actually plans and thinks.
Read original →