THE BRIEFING
Anthropic's Opus 4.7 quietly got more expensive to run despite flat list pricing—token counters reveal higher context-window costs. OpenAI released a lightweight multi-agent framework for Python, while TSMC's cautious earnings suggest the chip giant isn't betting big on sustained AI demand. Meanwhile, GitHub's fake-star economy and enterprise adoption stories paint a maturing, messier landscape.
Simon Willison's token counter tool exposes what Anthropic didn't advertise: Opus 4.7 consumes more tokens per request than 4.6, meaning higher real-world costs even though per-token pricing stayed flat. This is a classic cloud pricing sleight-of-hand, and developers are not amused. Watch for Anthropic to either clarify or face mounting pressure to revert the change.
OpenAI quietly dropped a native multi-agent orchestration framework for Python—trending hard on GitHub. This is OpenAI formalizing what the community has been hacking together with LangChain and AutoGPT. The fact that it's lightweight and opinionated suggests OpenAI sees agents as infrastructure, not science experiments. If you're building agents, this is now the reference implementation.
Ben Thompson flags that TSMC's earnings call was notably cautious—leadership isn't acting like they believe the AI chip boom is sustainable. They're building N3 fabs, but the tone suggests hedging, not conviction. If the world's chipmaker-in-chief is skeptical, that's a data point worth remembering when VCs pitch infinite GPU demand.
Anthropic updated Claude's system prompt between 4.6 and 4.7, and Willison's diff reveals subtle shifts in how Claude reasons and responds. Anthropic remains the only major lab publishing these prompts, which is both rare transparency and a gift to prompt engineers.
Hyatt rolled out ChatGPT Enterprise globally, using GPT-5.4 and Codex for operations and guest experience. This is OpenAI's playbook: land enterprise customers, tout productivity gains, and make the models feel indispensable.
Jack Clark's weekly roundup covers automated alignment research, a safety analysis of a Chinese model, and HiFloat4—a new quantization technique. Always dense, always worth skimming for what's moving in research.
Nvidia's Nemotron OCR v2 uses synthetic training data to achieve fast, multilingual OCR. Synthetic data is becoming the norm for specialized tasks where real-world labeled data is scarce or expensive.
Simon Willison upgraded his Claude token counter to compare token usage across models. This is how the Opus 4.7 pricing issue got exposed—tooling like this makes cloud pricing legible.
Thunderbird (yes, the email client) built an AI framework emphasizing user control and no vendor lock-in. This is open-source positioning against ChatGPT and Claude's walled gardens.
Manifest routes prompts across models to optimize cost and performance. As LLM pricing gets messier (see Opus 4.7), routing layers like this will become essential infrastructure.
An investigation exposes fake GitHub stars inflating AI repo credibility. This is the bot-farm problem coming to open source—trust signals are now gamed at scale.
DeepSeek released optimized FP8 matrix multiplication kernels. Fine-grained scaling improves inference efficiency—important as models scale and inference becomes the bottleneck.
A Model Context Protocol (MCP) server for indexing and searching codebases inside Claude. MCP is Anthropic's attempt to standardize agent-tool interfaces—watch this space.
Vercel's experiment in making localhost URLs agent-friendly. Small ergonomic wins like this matter when agents are your primary 'users.'
Another RAG framework enters the fray. The ecosystem is crowded, but adoption will come down to docs and real-world reliability, not GitHub stars.
Willison shares patterns for pulling Datasette queries directly into Google Sheets. This is the kind of unsexy glue work that makes AI-generated analytics actually useful.
Adobe is hedging against AI-native competitors by launching an enterprise agent platform. This is Adobe trying to stay relevant as generative AI eats its Creative Cloud moat.
Salesforce is positioning Agent Albert as proof that AI augments SaaS rather than replacing it. Wall Street is skeptical—this is Salesforce's existential pitch.
Fraud charges against iLearningEngines' former execs. The AI gold rush is attracting the usual grifters—this won't be the last indictment.
Google is designing custom AI chips with Marvell, aiming to reduce Nvidia dependence. Two million chips signals serious in-house infrastructure buildout.
The NSA is reportedly using Anthropic's Mythos model. This is Anthropic's FedRAMP play paying off—though the optics of 'AI for spies' won't help the safety narrative.
Anthropic's revenue is reportedly surging, and whispers of a trillion-dollar valuation are circulating. This is venture froth, but it reflects real adoption momentum.
A German court ruled an AI-generated comic based on a copyrighted photo doesn't infringe. This is transformative use doctrine meeting generative AI—precedent to watch.
Hundreds of AI influencers are pushing pro-Trump content on TikTok and Instagram ahead of midterms. This is the 2016 bot problem, now with photorealistic faces and GPT-4 captions.
Chinese humanoid robots beat human runners in Beijing's robot half marathon. This is mostly PR, but the pace of bipedal robotics progress in China is real.
Swiss government is pushing to reduce Microsoft dependence, citing sovereignty concerns. Expect more of this as AI gets baked into Office and Azure.
TRELLIS.2, an image-to-3D model, now runs locally on Apple Silicon. This is the prosumer generative 3D story accelerating—no cloud, no API keys.
A prompt-to-diagram demo running Gemma 4 entirely in-browser via WebAssembly. 3.1GB is hefty, but this is on-device inference crossing the usability threshold.
Supply constraints on high-bandwidth memory (HBM) could persist for years. This is the AI infrastructure bottleneck no one wants to talk about—GPUs are useless without RAM.
Historian Ada Palmer breaks down Francis Bacon's taxonomy of thinkers. Not explicitly AI, but useful framing for how we categorize LLM reasoning modes.
Bill McDermott explains how ServiceNow is tackling cybercrime at scale. Cybercrime is now the world's third-largest economy—bigger than most countries.
McDermott reflects on whether AI is eroding human connection. Short, philosophical clip—worth watching if you're tired of pure optimism.
90% of ServiceNow's customer service cases are now handled by agents. McDermott insists this is good news. You decide.
Jensen Huang explains Nvidia's venture strategy. This is how Nvidia is hedging its GPU dominance by becoming the default partner for every AI startup.
The All-In crew speculates on why Anthropic delayed Mythos' public release. The consensus: safety theater or competitive positioning, pick your poison.
The besties push back on AGI hype. Useful reality check amid the current capability inflation cycle.
A quick take arguing Mythos isn't as groundbreaking as Anthropic's marketing suggests. Worth a watch if you're skeptical of the hype.