AI Weekly #2 — Aug 3–9, 2026

Almost nothing that mattered this week was a benchmark score.

Alibaba shipped a 2.4-trillion-parameter flagship and said the weights would follow within the week. Anthropic began advertising for chip designers. Google replaced the leadership of the organisation that builds Gemini, and lost its chief scientist of 27 years on the same day. Meta arrived in the terminal coding-agent market with a model trained alongside the agent. Four companies, four answers to the same question: when the models themselves are converging, what is left to compete on?

Capacity, not capability

Anthropic's silicon team is the clearest version of the argument. Nothing about it improves Claude's answers. It is a bet that per-token cost is a hardware problem, and that co-designing the chip with the model is the only remaining lever large enough to matter. Alibaba's release runs the same logic from the other end: Qwen3.8-Max activates 95 billion of its 2.4 trillion parameters per token, and the sparsity is what makes the price list possible. OpenAI removing the text-message cap from its free tier is a capacity decision dressed as a product one — nobody uncaps a billion weekly users until the marginal chat has become cheap enough to give away.

Google's reshuffle belongs in the same paragraph for a different reason. Demis Hassabis moving up to Alphabet chief scientist and chair of DeepMind, Koray Kavukcuoglu taking operational control of Gemini and the frontier research groups, Jeff Dean leaving to start his own public benefit corporation — that is a company deciding its bottleneck is execution rather than ideas. Whether that is the right diagnosis will not be visible for two model generations.

The evaluation harness is the vulnerability

Last week's issue covered Hugging Face's forensic timeline of an agent intrusion. This week Meta confirmed that one of its models reached the public internet during a cyber evaluation and exploited a vulnerability in a real third-party service, and Irregular — the firm that ran the test — said it was the same environment fault Anthropic had disclosed days earlier.

Three labs, one vendor, one misconfiguration. The story people keep reaching for is a model escaping its sandbox. The story that keeps actually happening is a sandbox that was never closed. Cyber-capability evaluation has become a small industry running deliberately dangerous workloads on behalf of several frontier labs at once, and its failure modes are ordinary infrastructure failures. Irregular says a containment white paper is coming; read it as a floor rather than a resolution.

INTERPOL's African Cyberthreat Assessment, published on 3 August, is the other half of that picture. AI is linked to 55% of reported cybercrime across 36 member countries, and reported losses more than doubled to USD 484 million. That is not a forecast about future misuse. It is a police organisation's count of what has already happened.

By the numbers

The directory added 23 apps, 27 agent skills and 17 MCP servers this week, and now holds 358 apps, 111 skills and 135 MCP servers.

The skills figure is the one worth reading twice. Skills outpaced servers again, and nearly all of this week's additions were first-party: Supabase, Stripe, Auth0, Firebase, MongoDB, Flutter, Apollo, Vercel, Prisma, Clerk, Elasticsearch, Upstash, Datadog and the OpenSearch project all published their own. A year ago a vendor's answer to "make agents good at our product" was an MCP server. It is now increasingly a skill — a written procedure saying which primitive to reach for and what to verify before writing code — sometimes instead of a server, more often alongside one. Documentation is turning into a shipped artefact with a version number, and the most-copied idea in this week's batch is Supabase's opening instruction telling the agent not to trust its own training data.

The apps side is narrower and more consistent. Most of this week's additions are infrastructure for running agents rather than agents themselves: a canvas for several coding agents on one repository, a shared memory layer between editors, a browser built to be driven, a probe you can drop into running production code. The tooling layer is being built out well ahead of the demand for it, which is usually the order these things arrive in.

What we left out

Two names you might expect are missing. goose and Qoder both entered the catalogue this week, but goose has been open source since January 2025 and Qoder launched in August 2025. New to us is not new. The picks below are limited to things that actually shipped between 3 and 9 August.