Anthropic catches its breath, bots start paying to browse, and Garry Tan shows how to build a personal Wikipedia for your AI. Episode 9 connects infrastructure, economics and method.
Anthropic + SpaceX: a computational lifeline
The news of the week: the agreement between Anthropic and SpaceX to use the inference infrastructure of the Colossus supercomputer. The result is immediate: Anthropic has eliminated peak hours and has officially doubled its usage limits, stabilizing the workflow of thousands of teams. As a bonus, a reflection on the future of compute: bringing the models directly onto the chip.
GPT-5.5 Instant, Grok 4.3 and the Mythos effect
A week loaded with releases: GPT-5.5 Instant (faster and smoother), xAI’s Grok 4.3, and NVIDIA’s Nemotron 3 Nano Omni. But the most impressive data point comes from cybersecurity: thanks to Mythos, Mozilla has scaled from 30 to 420 monthly bugfixes in Firefox — the model finds and patches vulnerabilities at a speed that redefines what maintaining software means. As a curiosity, Higgsfield’s virality predictor maps brain reactions to anticipate a video’s success.
HTTP 402: the internet where bots pay
What if, instead of blocking bots, we charge them? That’s the proposal from Erik Voorhees: use the HTTP 402 protocol (“Payment Required”) so websites can request micropayments from agents. Tools like x402scan are already exploring that future where each agent has its own budget and pays per query. Free scraping’s days are numbered — or at least, it has a price now.
Coinbase and the “magic number”: one-person teams
Coinbase bets on one-person teams, eliminating middle managers to gain agility. It’s the organizational face of the agentic era: while Cloudflare and Coinbase itself restructure with massive layoffs, others like Linear grow progressively. In parallel, the chip war between Huawei and Nvidia intensifies, with China building its own infrastructure in the face of restrictions.
Gbrain and the “Skillify” habit
The closing is methodological and very practical. First idea: the model on its own is “dumb” — the value is in giving it structure through skills, memory and rules. Second: Garry Tan’s Gbrain project, an architecture for creating a “personal Wikipedia” in Markdown that gives AI all your context. And third, the “Skillify” habit: turning every repetitive process into a deterministic skill the model runs on its own. Document once, delegate forever.
Thanks for joining us one more week on Tokenizados. Subscribe and we’ll talk next episode.