Another Daily AI Newsletter - July 28
⭐ Top Story: Dario Amodei challenges the open-weight letter
A broad group of technology companies published a letter supporting open-weight AI, arguing that access to model weights expands competition, gives customers more control, and lets outside researchers inspect systems for weaknesses. Anthropic did not join the letter.
Dario Amodei responded with Anthropic’s full position. He said Anthropic has never advocated for a categorical ban and called open models without dangerous capabilities a public good. His disagreement is narrower and more consequential: he does not believe open access necessarily makes safeguards easier to build or gives defenders a greater advantage than attackers.
The industry’s evidence is still split. The Open Secure AI Alliance points to the response to the recent Hugging Face incident. Closed models reportedly refused some forensic work, while an open model helped investigators inspect more than 17,000 recorded actions. The alliance presents that case as evidence that inspectable models can be valuable defensive tools.
Anthropic focuses on what happens when a model becomes capable enough to assist cyberattacks or biological attacks. Open weights can be modified to remove safeguards, used without monitoring, and cannot be recalled after release. Amodei wants policy tied to capability: tighter controls on advanced chips, action against industrial-scale model distillation, and mandatory safety testing for sufficiently capable models whether their weights are open or closed.
The position drew criticism from the open-model community. AI researcher Nathan Lambert called it a reasonable restatement of Anthropic’s views but rejected the proposed restrictions on distillation. Investor Sheel Mohnot described the capability-based approach as principled while noting that it could also protect the advantage of closed-model companies.
The debate now turns on a question that neither side has settled: when a powerful model is released openly, do more capable defenders gain the larger advantage, or do attackers?
Cyber defense is becoming a coordinated agent system
Microsoft introduced Project Perception, MAI-Cyber-1-Flash, and MDASH. Project Perception coordinates red-team agents that search for attack paths, blue-team agents that investigate risk, and green-team agents that apply fixes. Microsoft says its MDASH vulnerability system, using the new specialized model, scored 96% on CyberGym while costing almost 50% less than its current configuration. Those are Microsoft benchmark results; Project Perception enters public preview on August 3.
Vercel’s DeepsecBench shows why a system still needs multiple models and human review. The benchmark uses 231 human-judged security findings. Its best-tested configuration found 30.7% of the known vulnerabilities with 96.3% precision, while lower-cost models produced different tradeoffs between recall, accuracy, speed, and price.
The security product is becoming a coordinated loop: specialized agents search, investigate, and repair; benchmarks measure what they miss; people remain responsible for validating the result.
Robots are getting longer jobs and more human teachers
Enigma raised $71 million and put more than 100 real robots online for anyone to control from a browser. The startup is testing whether people prefer to guide robots with text, speech, demonstrations, or direct manipulation. Those interactions will also produce real-world data that Enigma hopes can improve its robot models.
The new τ₀-VLA system plans and executes household tasks lasting as long as 12 minutes. It breaks work into subtasks, predicts possible outcomes before committing, and updates its memory when the physical result does not match the plan. Across four real-world tasks, the researchers report a 45% average success rate with hierarchical planning versus 27.5% for direct execution using the same underlying action policy.
Nvidia says its Cosmos world models have passed 10 million downloads. The number does not reveal how many systems are in production, but it shows how quickly shared models and data for robotics and autonomous vehicles can spread through the developer ecosystem.
These projects attack the same bottleneck from different directions: better planning helps robots stay on task, while broad human interaction supplies the examples needed to make their behavior easier to guide.
AI is widening jobs before companies rewrite them
OpenAI studied more than 800,000 messages from U.S. ChatGPT users and found people using AI outside the traditional boundaries of their roles. Across work-related messages, 16.8% involved a task associated with another occupation. After common activities such as writing and scheduling were excluded, 43.5% of occupation-specific messages crossed job boundaries. Designers used AI for engineering and marketing work; salespeople analyzed data; small-business employees took on tasks that might otherwise require a specialist.
Cohere launched North Automations so employees can describe workflows in ordinary language. The product is designed to let business teams build repeatable AI processes without becoming software developers.
The early change is visible in the task list. AI gives the person closest to a problem a way to attempt more of the surrounding work before handing it to another department.
Quick Hits
Safe Superintelligence and Nvidia announced a long-term partnership built around Vera Rubin systems. SSI says the agreement and Nvidia’s investment will increase its compute by an order of magnitude. The official announcement does not disclose the investment; TechCrunch reports that Bloomberg placed it at $5 billion.
Google’s AI Overviews now appear in 43% of searches, according to Similarweb data reported by TechCrunch. That is up from 15% a year earlier, while visits to Google’s conversational AI Mode more than doubled between June 2025 and May 2026.
Public Claude share links and Artifacts appeared in Google search results. Anthropic says privately sent links were not discoverable unless someone posted them publicly. The results had disappeared from Google by Monday afternoon; Claude users can review public links under Settings, Privacy, Shared Chats.
Satya Nadella warned companies against handing one model provider their context, memory, and workflow layer. He argued that businesses should preserve their own usage metadata and keep the agent harness separate enough to switch among models.
Meta added Muse Spark, Threads, and neural handwriting to its display glasses. The update brings a new voice model, lets wearers listen and respond to Threads posts, and begins early access for writing short messages by tracing letters with a finger.
GPT-Live is now available in eligible ChatGPT Business, Enterprise, and Edu workspaces. Voice in Chat supports more natural interruptions and follow-up questions, while Voice in Work and Codex can start tasks and coordinate agents from the desktop app.
Kimi K3 arrived in Cursor and Notion through U.S.-based inference providers. Cursor says it supports zero-data-retention inference; Notion says its deployment is hosted by Fireworks.
🔬 Research Radar
Nvidia released Ising Calibration 1.5 to automate the diagnosis and tuning of quantum processors. The 31-billion-parameter vision-language model can interpret unfamiliar calibration plots and recommend next steps. Nvidia released weights, quantized versions, data, evaluation tools, and deployment blueprints; its performance comparisons remain company-reported.
Qwen researchers introduced Skill Self-Play, a training loop in which models generate, solve, and verify new tasks. A proposer creates challenges, a solver attempts them, and a controller expands the library of skills used to keep the tasks varied and checkable. The team also released its code.
Researchers from Stanford, Google, and several leading AI labs proposed task-based tests for medical AI “superintelligence.” Their argument is that broad benchmark scores do not establish clinical capability. Evaluation should be tied to defined medical tasks, realistic operating conditions, and evidence that performance translates into better care.
🛠️ For Builders
Vercel added WebSocket mode for the OpenAI Responses API to AI Gateway. A persistent connection avoids repeatedly sending the full conversation state. Vercel reports roughly 40% lower end-to-end latency in workflows with more than 20 tool calls.
Cloudflare added `createTestHarness()` to Wrangler. It runs production Worker builds locally, supports multiple Workers, can mock outbound requests, and works with Node.js test runners and Playwright.
GitHub extended enterprise-managed settings to the Copilot app and cloud agent. Administrators can centrally control approved plugins, marketplaces, model defaults, and whether developers may bypass command and network approvals.
Nvidia released NOOA, an object-oriented research harness for AI agents. Developers define agents as Python classes with typed state, tools, memory, and methods, making the reasoning loop easier to inspect and modify.
AWS published an open implementation of task-aware knowledge compression. It pre-compresses document collections differently for finance, legal, or other tasks, then routes each question to an 8x, 16x, 32x, or 64x context tier.
📘 AI Term of the Day: Checkpoint
Google defines a checkpoint as data that captures the state of a model’s parameters during training or after training is complete.
Think of it as a saved version of what the model has learned. A training team can stop, reload the checkpoint later, and continue from that point instead of starting over. The model parameters stored in a checkpoint are also the central artifact shared in many open-weight releases, which is why the term matters in today’s debate.
Google’s definition | Go deeper with Hugging Face’s guide to saving and loading checkpoints