I embed with companies and ship the AI production systems they actually run on.
Paris
Joined June 2007
- Tweets1.2K
- Following872
- Followers878
- Likes2.3K
Pinned Tweet
I don’t want one giant context soup.
Honcho is short-horizon working memory. I built Hypermnesic for the other half: long-term, markdown in git, files are truth, the index is disposable.
LongMemEval V1, GPT-4o-judge axis: on par with Mastra, +12 vs Zep, +23 vs no-memory. Retrieval recall@10 is 0.949.
uv tool install hypermnesic hypermnesic local-proof --demo-dir /tmp/hypermnesic-demo
I wouldn't have bet a bank would self-host open weights this early.
The tools did not move: Claude Code and Codex, with the model swapped behind them. The rationing stopped. At about 30 cents per million tokens, AI went from hard tickets to everything around the code, including the inbox and Jira.
When tokens stop being the scarce input, supervision becomes the constraint.
BREAKING: Revolut moved ~2,000 engineers onto self-hosted GLM for coding and pushed Claude Code and Codex to the exception list.
Per David Abitbol, who leads AI ops there: 56 Nvidia B200s, a blended 30 cents per million tokens, roughly a tenth of what they paid Anthropic and OpenAI. Engineering lead time improved 70%.
Here's the part that matters: the gains showed up when tokens stopped being rationed. Engineers quit saving AI for the hard tickets and pointed it at everything.
Frontier models become the exception, open weights become the default. That's a pricing problem for every lab heading to IPO.
h/t @OliverMolander
Would you move your team off frontier APIs if the bill dropped 90%?
1
51
I've used Cadu, the unofficial Hermes iOS client, for months. It put the phone's contacts, health, location, reminders inside the agent. Hiring its builder @fl_rn_st was the best call @NousResearch made. He says the core stays similar as Cadu goes official. What lands first?
1
85
cua is one of the most underrated companies in the AI space
1/ We're reimagining what it means for your agents to work with all your computers.
Today we're excited to share Cua Spaces, built on Cua Driver and our virtualization stack, rolling out on macOS today.
Download it now, it's free and source-available: spaces.cua.ai
Zero dependencies: the runtime imports nothing outside the Python standard library. One SQLite file, 0600, in my own data directory. Self-hosted, no platform account, AGPL-3.0-or-later (4/5)
1
12
I use it in my own work. pip install paraphe, and there is a 15-second console demo on the README.
github.com/leonardsellem/par…
(5/5)
1
14
deepseek-v4.1-flash is out. First gut: insanely fast. I can't even keep up reading its reasoning.
Crazy cheap, and peak-hours pricing lines up with Europe work hours. Looks quite verbose. Will that offset the cheap cost per million tokens?
Deepseek claims only Astra can beat it. So far, so good. I haven't shipped code to production with it yet.
1
1
65
fwiw @FrameworkPuter wants to save the world from hardware obsolescence, but can't cancel an order placed hours earlier.
Their answer: let the laptop cross the Atlantic twice, then return it. Reason given, their systems "transmit one way" to their logistics partner.
1
44
Si vous ne devez lire qu'un article sur le release dont tout le monde parle. Merci @UlrichRozier 🙏
Omarchy est le projet le plus discuté du moment dans l'écosystème Linux : 100 000 téléchargements en une semaine, 10 M$ sur la table, signé DHH (Ruby on Rails)
J'ai voulu comprendre d'où vient l'engouement, et surtout s'il est justifié. Verdict après test sur @frandroid frandroid.com/culture-tech/3… et j'ai fait un guide avec mes erreurs à ne pas reproduire chez vous omarchy.ulrichrozier.com/ @dhh @OmarchyLinux
1
77
Found Paseo yesterday. Never heard of it. Open source, still beta, more polished than most VC-funded IDEs I have been using for months.
One session fans out Claude, Codex, Grok, Hermes. Claude Code on my Max, no extra Anthropic API bill.
paseo.sh
92
I haven't touched my weekly Claude Code quota. Ox Alpha and Grok 4.6 did the work. I started because it was free. I stayed because it was good. GPT-5.6 only when the plan got hard.
Introducing GLM-5.3-Flash
- Leading capabilities at a highly competitive price
- Natively multimodal with a 1M-token context window
- A 320B-A18B model released under the MIT License
- Previously previewed as Ox Alpha, running entirely on Chinese AI chips
Blog: z.ai/blog/glm-5.3-flash
Available now across all official platforms:
Weights: huggingface.co/zai-org/GLM-5…
API: docs.z.ai/guides/llm/glm-5.3…
Coding Plan: z.ai/subscribe
ZCode: zcode.z.ai/en
Chat: chat.z.ai
AutoClaw: autoclaw.z.ai
65
I accept downtime. I do not accept Actions failing five times in August. @cursor_ai Origin and @jack's Buzz are shipping forges. The leftover job is still CI. There is no locked position.
22
Vercel didn't open-source a tool. They published the company loop.
Once other teams run fx, they inherit who signs off.
37
