FluxDigital

New Anthropic research: A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude.

It is a 2 to 4T param model. They are serving it across 70-100 wafers. To get healthy serving characteristics, they are essentially putting at most one layer per wafer, and the model is in the ballpark of 70-90 layers. There's a couple of different ways this could be served and model sizes implied by that. One is if they keep the heavy KV caches they've used before. Another is if they go with lighter KV cache designs more akin to DeepSeekV4 or Hybrid SSM models. The fact that they've partnered with Cerebras and designed with the hardware in mind means they're much more likely to have gone the second route. That SRAM bandwidth is too precious for a heavy KV cache. As such, something like the below is the center of probability mass: 3T total, 150B active, 70 layers.

Another major Grok Build update just landed, packed with new features, extensive bug fixes, and meaningful performance improvements Release Notes: v0.2.84 — 2026-07-03 Features: • Announcements now update live during active sessions without restart or /new. • Hiding an announcement no longer suppresses later criticals; new ones reappear automatically. • run_terminal_cmd now requires a one-sentence description rationale in every invocation. • ask_user_question timeout policy is now configurable in config.toml and /settings. • Ask-Question timeout can now be toggled from /settings (Agent & Approval). • Thinking/reasoning blocks are now shown by default while the model is working. • Critical announcements now show a red title with a clickable [hide] button and aligned message. • Added remote_fetch option under [features] in config.toml to disable all backend catalog and settings fetches for air-gapped environments. Bug Fixes: • Images pasted or read from GIF, BMP or TIFF files are now automatically converted so they work with image generation. • Queue panel now shows action buttons on hover and the status bar displays a compact done/total task count. • Hook matchers now correctly see the real MCP tool name instead of the internal dispatcher name. • Copy now succeeds when running inside containers even when the terminal brand cannot be detected. • Tool result previews no longer paint opaque panels in grok --minimal. • grok wrap now correctly handles quoted strings and shell aliases. • Text selection settings now correctly honor explicit keep_text_selection values even when legacy keys remain. • Fixed a freeze that could occur when editing and sending the last message in the queue. • Fixed a startup crash on minimal Linux systems lacking system CA certificates. Performance: • Grep now stops early on broad searches, returning faster results with far less memory use. • Idle CPU and memory usage after long sessions or resume is now dramatically lower.

Precast concrete operation at TBC Dubai. The first phase of Dubai Loop will utilize ~25,000 segments, each weighing ~4,000 pounds. Total tunnel weight: ~100 million pounds!

𝕏 is the group chat of humanity