Sethian (@theSethian)

GODFATHER OF AI ON PBS LAST NIGHT: "HOW CAN I SAY THIS... MAYBE A CORRUPT SENILE FELON IN THE WHITE HOUSE." HE THINKS CHINA'S POLITBURO GETS THE AI RISK BETTER THAN THE US, AND THE TWO WON'T TEAM UP FOR 2 YEARS. On AI, Geoffrey Hinton is someone you can trust. He won the 2024 Nobel Prize in Physics for the research behind today's AI. Last summer he went to Shanghai and talked to one of Xi's protégés at the top of the Communist Party. His words on Firing Line: "I believe the Politburo understands this stuff much better than the U.S. government." Hinton wants Washington and Beijing to work together so superintelligence doesn't take over. Then he explained why that has to wait: "It won't happen while we still have a, how can I say this, maybe a corrupt senile felon in the White House. So it's not going to happen for two years." The host grinned. Do you agree with the Godfather of AI?

DeepSeek spent 88 minutes building a horror game. One sound made it worth sharing. It includes a monster, flashlight batteries, and a three-key escape. When the creature gets closer, the heartbeat accelerates. That single mechanic communicates the threat without a tutorial. The creator flinches before he finishes explaining it. Nobody sends a friend a feature list. They send the moment that made them feel something. Kai's article below offers five buildable game shapes that can each be explained in one sentence. People will forget the 88-minute build. They will remember when the heartbeat changed.

GLM-5.2 built a working Chrome extension in 10 minutes. Matt Wolfe asked it to create PageBrief with: > current-page summaries > extracted action items > important links > Markdown export > a clean popup > every file needed to install it GLM returned the folder. Wolfe opened Chrome, enabled Developer Mode, clicked "Load unpacked," and selected the generated files. Then he tested PageBrief on an article. The extension produced a summary, key takeaways, important links, and clean Markdown ready to copy. The extension proves GLM-5.2 can handle one small, useful build. Beyond that demo, it has MIT-licensed open weights and a 1M-token context window. On reported benchmarks, it scores 62.1 on SWE-bench Pro, ahead of GPT-5.5's 58.6. On Terminal-Bench 2.1, it reaches 81.0, a few points behind Opus 4.8 at 85.0. The API costs $1.40 per million input tokens and $4.40 per million output tokens. In the article below, routine fixes, frontend work, tests, refactors, and long runs go to GLM. Opus keeps the hardest repo-level and security-sensitive work. GLM ships the extension. Opus can wait for the code that actually needs it.

One sentence gave Claude Code a researcher, a writer, and a reviewer. The guy in the video installs CrewAI's four official Claude Code skills and types: "build me a research crew" Claude Code installs CrewAI, scaffolds the project, creates each agent with its own role, goal, tools, and memory, then pulls CrewAI's live docs through Ask Docs while it builds. When he runs it: > the researcher gathers the material > the writer turns it into a draft > the reviewer checks the result > each agent passes its output to the next No manual handoff between three chats. At the end, he runs multiple Claude Code sessions through DevSwarm, each on its own Git branch, so separate agents can work in parallel without editing the same branch. The demo ends with a working crew. The article below adds the pieces needed to keep that crew useful on real software: > Task Master turns the PRD into tasks and dependencies > specialist agents isolate research, implementation, testing, and security > Claude-Mem carries verified decisions across sessions > ccusage exposes the token bill > Claude Code Router sends cheap work to cheaper models > Ruflo coordinates jobs that can genuinely run in parallel The repositories cost $0 to install. The model calls don't. If five agents return code you delete, the terminal was busy but the system wasn't working.

GPT-5.6 Sol Ultra built a playable Geometry Dash-style game in one prompt. Prompt Potato keeps raising the difficulty for 12 minutes: Geometry Dash, Doom, then Rocket League. At 00:18, the first build arrives with five difficulty levels, different movement forms, gravity changes, portals, music for each level, an icon kit, and a progress counter. It looks polished, but the hardest level has a real input problem. Jumps don't always register after landing, so he spends 700 attempts on it before enabling invincibility to see the ending. At 04:14, Sol builds the Doom-style game. The first enemy deletes the player in seconds because several rifle attacks stack into one burst. Sol finds the damage issue and repairs it. The next run has usable weapons, a keycard, multiple enemy types, and a full boss fight. Rocket League is the hardest test: > 1v1 and 3v3 modes > a garage with car colors and boost trails > goal replays > air rolls, aerials, flips, and demolitions > bots that can save shots and also park themselves on the wall At 08:59, one correction fixes the air roll camera and flips, adds free play, and gets the bots moving again. Their intelligence still needs work, but the aerial controls and even a tornado flick work. This is the kind of job that can justify Sol: a long build with several systems, visual judgment, debugging, and multiple rounds of repair. The expensive mistake is letting a demo like this choose the model for every other task. Sol costs five times as much as Luna: > Sol: $5 input / $30 output per million tokens > Terra: $2.50 input / $15 output > Luna: $1 input / $6 output The article gives each tier a specific job: > Luna for classification, extraction, tagging, and simple summaries > Terra for chat, RAG, drafting, code review, and standard agent steps > Sol for long builds and the cases where Terra fails your eval > Ultra only when parallel agents earn the extra token spend > retry on the next tier only after validation fails > log the model and token count on every call Sol earned its place on these game builds, and it doesn't need to classify the support ticket.

You're paying Claude to read garbage it never uses. Headroom cut one code search from 17,765 tokens to 1,408 and returned the same answer. Nathan Hodgson's video shows Headroom at 39K GitHub stars. The repo has already passed 59K. Claude Code carries logs, files, database dumps, and tool results into context. A few useful lines can arrive wrapped in thousands of tokens of noise. Headroom sits between the agent and the model, compresses that clutter before Claude reads it, and passes on a smaller version. Installation takes one command. The 92% figure belongs to that specific code-search benchmark. It isn't a guaranteed discount across every session. The article below targets terminal output with RTK, using a Claude Code hook that rewrites Bash commands before they run: > git status becomes rtk git status > tests return failures instead of every passing line > repeated log messages are grouped > builds, diffs, and search results are condensed In RTK's 30-minute project estimate, command-output usage falls from roughly 118,000 tokens to 23,900. RTK only intercepts Bash calls. Claude's built-in Read, Grep, and Glob tools bypass the hook, so this works best when terminal output is eating the context. If Claude spends most of the session in Bash, I'd start with the RTK setup in the article.

NVIDIA's $4,600 personal AI supercomputer loses the speed round to four used RTX 3090s. Tech2WiLD owns both setups and has been running them for nearly two months. At 00:31, he compares the cost. One DGX Spark sells for roughly $4,000 to $4,600. He paid about $3,500 for three used RTX 3090s and already owned the fourth, plus a compatible motherboard. Two of those 3090s run Qwen 3.6 35B-A3B at up to 150 tokens per second on one stream. The full rig has 96GB of VRAM. One Spark has 128GB of unified memory. The 3090 rig is faster for local models that fit. Spark's extra 32GB leaves more room for context, parallel agents, or another model, but one Spark still runs roughly the same class of models. At 04:52, the 3090 build hits its limit. His consumer motherboard is already full at four GPUs. Moving to a Threadripper-class platform can cost another $4,000 to $5,000 before buying the next GPU. A second Spark adds another 128GB and opens larger models such as DeepSeek V4 Flash and MiniMax M3 without rebuilding the whole machine. His verdict at 10:28 the 3090 rig for speed right now, the Spark if the setup is going to keep growing. The article below takes the other side. It compares Spark's 273 GB/s memory bandwidth with 936 GB/s on one RTX 3090, then ranks 19 GPUs by bandwidth. For one box, the 3090s win. Spark starts making sense when the next box is already part of the plan.

Installing every Claude plugin can make Claude Code worse. Tech With Tim opens with the failure mode, then spends 22 minutes cutting the stack down to tools with a specific job: > 00:00 once Claude sees around 50 tools, it starts picking the wrong ones > 03:06 Pyright checks generated Python against real type errors > 05:06 Anthropic's GitHub plugin fails; the MCP workaround is connected by 07:54 > 16:21 Context7 pulls current framework documentation > 17:34 Composio finds the required tool on demand instead of loading the full catalog into context > 19:59 Figma gives Claude the source design before it writes the page The article below adds Playwright for browser checks, Postgres or Supabase for data, Slack for team updates, and Gmail, Linear, or Notion for the work around the code. One workflow starts with a failed CI check on a pull request. Claude reads the failure, queries the database to reproduce the bug, writes the fix, and posts the result in Slack. The author keeps the setup to four to six MCP servers, with read-only database access and least-privilege tokens. Keep the stack small enough that Claude can still choose the right tool.