A.I & Blockchain Engineer
Joined February 2008
- Tweets2.4K
- Following3.1K
- Followers473
- Likes4.5K
Pinned Tweet
gave openai dots a week with a real workflow and moved everything back to grok bot after 7 days. one dot only, a vm i couldn't use as a human, and zero way to hand it a credential. long write-up of what broke for me
chebyte retweeted
1/ Introducing the Cua Spaces SDK.
Every big lab now ships a bot with its own computer: Grok Bot, Muse, dots. None let you bring your own. With Cua Spaces, now you can.
Put Claude Code, Codex or any agent on compute you choose, and run jobs across all your computers at once.
chebyte retweeted
OMG DeepSeek V4.1-Flash-1017 Search & Fetch APIs costs
$0.75 per 1000 calls.
OAI & Anthropic: $10 per 1000 calls.
Grok: $5 per 1000 calls.
Google Gemini: $14 per 1000 calls.
OMG 😧
🔍 Search & Fetch APIs are now open to everyone!
Powered by our in-house search engine, serving hundreds of millions of requests daily worldwide.
💰 Each endpoint: $0.75 per 1,000 calls.
📖 Docs: api-docs.deepseek.com/guides…
📌 Search & Fetch calls via server-side tools in DeepSeek Harness, Claude Code, and Codex are currently free for DeepSeek API users. Billing starts Oct 31, 2026, at 16:00 UTC.
3/4
chebyte retweeted
Prompt of the day:
Create a pre-commit hook that prevents files over 5,000 tokens. Tokens = bytes/4.
Run only on staged files. Ignore files which are not authored by agents, like lockfiles or codegen.
Provide explicit instructions on how to split files in the error message.
The simplest investment I ever made in my codebase. So good.
4 repos to get models off someone else's server:
1. ggml-org/llama.cpp, the C++ core a lot of local tools wrap
2. ollama/ollama, one binary, pull and run
3. ml-explore/mlx on Apple silicon
4. google-ai-edge/gallery, Gemma offline on Android, iOS, Mac
4 even shows its reasoning
Anthropic's numbers on its own engineering org:
→ ~6 agent hours per human hour
→ commits per engineer 5x in six months
→ junior technical hiring 27% to 9%
→ sessions wait on a person ~80% of the time
that 80% is the number i'd work on before adding a single extra agent
OpenCode got breached last night.
→ 1.4% of OpenCode Console users hit, only accounts with external providers connected
→ they tie it to a large scale distillation attack
→ affected users got an email this morning
if you ever pasted an OpenAI or Anthropic key in there i'd rotate it anyway, email or not
Android's team ships official agent skills, plain SKILL.md files.
→ XML to Compose, Play Billing, AGP 9, R8
→ android skills add --all wires them into your agents
written only where their evals showed models failing. i trust that over prompt packs
github.com/android/skills
chebyte retweeted
🐘 Introducing Postgres Internals in 3D.
🌟 pgrun.dev/pi
Stop reading boring diagrams. Start exploring PostgreSQL.
An interactive 3D map to understand how Postgres actually works.
100% free. Built for developers.
xitter.cf/alxshp/status/21100494…
Introducing PGSync 🐘
Real-time PostgreSQL replication. FREE. OSS
Built in Rust. 🦀
Postgres 16 / 17 / 18.
No extensions. No triggers. No complex setup.
→ snapshot
→ replicate
→ stream changes
→ track WAL
→ recover from crashes
→ resume automatically
→ keep databases in sync
~15 MB RAM during streaming.
Built for developers. Built for AI agents.
Open source.
xitter.cf/alxshp/status/21078790…
Claude Code now gzips the conversation it sends each turn.
→ first token, local: 4.07s to 1.33s
→ cloud: 16.99s to 6s, sessions boot while you type
→ opt out: CLAUDE_CODE_GZIP_REQUEST_BODIES=0
a 3x from compressing JSON, funny nobody did it sooner
code.claude.com/docs/en/chan…
paste this into CLAUDE.md before an agent touches tests:
## done
1. new test fails on main first
2. fix src/, never the test
3. test looks wrong? stop and tell me the line
4. paste the exact command + output
if the diff edits an assertion to go green i revert the whole thing
AgentPlans radar for Oct 17. 849 repos scanned, ranked by stars/day (no 7-day history yet):
→ deepseek-harness
→ morluto/rea
→ ponytail
→ affaan-m/ECC
→ dsh-our-free-model
All five MIT. No model weights on the list, just the layer around the model.
agentplans.vercel.app/radar
chebyte retweeted
New from Vercel Labs: tsc-ts
TypeScript was written in TypeScript
Then rewritten in Go
We ported it back to TypeScript
Compiled it with scriptc
And it's 2-4x faster than tsgo
Cline v4.1.25: tasks keep running after you close the editor. pick a repo + branch, it runs on their hosted box and pings you when it's done.
also in there: with no model pinned, the Cline provider now defaults to Gemini 4 Argon. if you didn't pin one, you're on a new model
NVIDIA open-sourced ASPIRE. the "trained model" is a repo of robot skills, not weights.
→ agent writes control code, reads the failure trace, patches it
→ 99% on 80 LIBERO-Pro tasks
README says the generated Python gets full imports. read that first
github.com/NVlabs/ASPIRE
The NIST P-256 seed people side-eyed for 25 years came out of lrand48, seeded with a Unix timestamp from August 1997.
→ Claude traced it in under 30 min, after a few Anthropic folks had chased the "hashed phrase" rumor and gotten nowhere
→ P-192 through P-521 all fit 1997-98 timestamps
B-163 and a few base points are still unexplained, so the backdoor crowd keeps a little room
anthropic.com/research/ecdsa…
Claude Mythos 5 ran in a loop against a circuit checker, no human between rounds, and beat the best published quantum circuits on 7 of 8 tasks.
→ elliptic-curve circuit: 17x fewer gates
→ 53h longest run
the scorer is what i'd steal: correctness first, then qubits x gates
chebyte retweeted
Introducing Supabase Middleware.
Write per-request logic once, like checking who's calling, and run it on any runtime that has fetch. The same middleware also works inside Hono, H3, Elysia, NestJS, and TanStack Start with a bridge you copy.
Learn More ↓
supabase.com/blog/supabase-m…
