Overview
Today had two clear threads: the internals of AI systems got more concrete (from Anthropic’s “global workspace” work to fresh arguments about cheaper memory for inference), and the product surface kept expanding, with new models, business integrations, and “just let the tool do the boring bit” workflows. In the background, there’s also the geopolitics of model access starting to look less theoretical.
The big picture
AI is being pulled in three directions at once. Researchers want to peer inside models and name the parts that do the hard thinking. Builders want cheaper, more open building blocks, whether that’s a 295B MoE under Apache 2.0 or a terminal library that makes everything snappier. Policymakers want control, and “who gets access to which model” is turning into a live question, not a future one.
Anthropic maps a “global workspace” inside Claude
Anthropic says it’s found something like a broadcast hub inside Claude, a set of activations that carry concepts around for silent, multi-step reasoning. The interesting bit is the split: remove that space and you don’t just get a dumber chatbot, you get fluent output that loses its grip on harder reasoning chains.
It’s the sort of work that makes interpretability feel less like vibes and more like instrumentation, even if it stops well short of any claim about subjective experience.
Access consciousness, not the spooky kind
Boris Power put a clean frame around the Anthropic result: it looks closer to “access consciousness” (stuff a system can use, report, and manipulate) than any test of phenomenal consciousness. That distinction matters because people keep mixing the two in everyday arguments about whether models are “awake”.
Also worth noting: the demo culture around this is growing, which is how ideas like “J-space” spread beyond a single paper.
Tencent drops Hy3, a 295B MoE with an open licence
Tencent Hunyuan announced Hy3: 295B parameters, mixture-of-experts, Apache 2.0, weights on Hugging Face, and a two-week free API push. The positioning is blunt: “rivals trillion-scale flagships” while staying usable for agent-style work without eye-watering bills.
Whether the benchmarks hold up in the messy real world is always the question, but the licensing choice is the headline for anyone building commercial tools on open weights.
Carmack’s pitch: model weights don’t need “random access memory”
John Carmack argued that inference has predictable memory access patterns, so we might not need expensive HBM for weights at all. His provocation is to treat flash as a serious option, accept cold-start pain, and design systems around sequential streaming, scratchpads, and careful caching.
If that direction pans out, it’s not a minor optimisation, it’s a cost curve change that decides who can run big models and where.
Open model benchmarking turns into a public sport
Chamath Palihapitiya says 8090 has had inbound from major open model providers and cloud platforms, and now plans “recursive optimisation” tests across Chinese open-weight models and American open source, starting with Nvidia’s Nemotron.
There’s a quiet subtext here: if open stacks can be measured, compared, and tuned like this, procurement starts to look more like engineering and less like picking a brand.
China weighs restricting overseas access to advanced models
Jukan summarised Reuters reporting that China’s Ministry of Commerce has been meeting with major AI companies about limiting overseas access to cutting-edge models, including unreleased systems and potentially open-weight releases.
Put next to recent US controls, it reads like an accelerating pattern: model capability is being treated as a strategic export, not just a product feature.
Micropayments for the web, built for humans and bots
skooookum pointed to Cloudflare’s Monetisation Gateway and the x402 protocol idea: a 402 Payment Required response that can trigger tiny payments without the usual account creation and redirect dance.
If that becomes normal, it changes how content, APIs, and datasets get sold, and it also changes how agents navigate the internet, paying as they go.
ChatGPT for PowerPoint goes worldwide
ChatGPT’s PowerPoint integration is now generally available across all plans. The appeal is simple: take the drudge out of turning notes into a coherent deck, while keeping the slides editable so you are not stuck with a black box export.
It’s another step in “AI inside the tools people already live in”, which tends to matter more than flashy demos.
Ghostty claims the terminal speed crown
Mitchell Hashimoto says Ghostty’s nightly build is now more than 2x faster than other “fast” terminals on IO throughput tests, and that the gains land in libghostty, so downstream users benefit too.
There’s a small cost (more memory, an extra thread), but if you spend your day in a terminal, responsiveness is not a luxury.
SpaceX lines up another rideshare swarm
SpaceX posted the Transporter-17 Falcon 9 launch stream. These missions are becoming their own genre: huge batches of small satellites, frequent cadence, and an ongoing argument about how crowded low Earth orbit should get.
The spectacle is real, but so is the need for tracking, deorbit compliance, and grown-up debris maths.
