It's only 150 million parameters (RoberTa derivative) and knowledge cut-off 01/01/2020, but it looks very promising if we manage to do that with much larger models. It's true that it's magical to watch the message appear instantly.
Feels a bit like GPT-2 in terms of intelligence.
Feels a bit like GPT-2 in terms of intelligence.
0
0
0
X is so vibecoded that I got redirected to a profile called @Settings when I clicked on settings 🤣
1
0
0
@thsottiaux You must be shitting me, my weekly was sitting at 95% left since the release of Opus 5.5 anyway
Will go with Claude and Pi+zai for now I guess
Will go with Claude and Pi+zai for now I guess
0
0
0
Numbers, from replaying all 11,267 API calls:
- average context at a rewrite: 563k tokens
- 1-hour cache on only the 18 agents that waited: -25.6%
- the agent in the image: 15 hours, 54 rewrites, 52 avoidable, half its cost
Details:
- the agent file needs Claude Code v2.1.248+, omptCacheTtl or CLAUDE_CODE_SUBAGENT_PROMPT_CACHE_TTL overrides it
- the 1-hour cache is ignored while a subscription is on usage credits
- on an API key, your main session is on 5 minutes too; set promptCacheTtl: "1h" for it
- replay method: API-price costs from every call's usage data; after a 5-60 min gap the prefix is read instead of rewritten, other writes cost 2x instead of 1.25x, gaps over 60 min still expire
Docs: code.claude.com/docs/en/prompt…
File: gist.github.com/shidenkai0/0e7…
- average context at a rewrite: 563k tokens
- 1-hour cache on only the 18 agents that waited: -25.6%
- the agent in the image: 15 hours, 54 rewrites, 52 avoidable, half its cost
Details:
- the agent file needs Claude Code v2.1.248+, omptCacheTtl or CLAUDE_CODE_SUBAGENT_PROMPT_CACHE_TTL overrides it
- the 1-hour cache is ignored while a subscription is on usage credits
- on an API key, your main session is on 5 minutes too; set promptCacheTtl: "1h" for it
- replay method: API-price costs from every call's usage data; after a 5-60 min gap the prefix is read instead of rewritten, other writes cost 2x instead of 1.25x, gaps over 60 min still expire
Docs: code.claude.com/docs/en/prompt…
File: gist.github.com/shidenkai0/0e7…
0
0
0
One agent file cuts Claude Code usage by 25% on a feature I built with 72 agents over 64 hours. Computed call by call on the real trace.
Unlike the main convo, sub-agents only keep their prompt cache for 5 minutes, even on a subscription. If one waits longer than that on a test t call pays for its whole context again. In that build it happened 200 times: 31% of the total cost.
The worst was one agent polling a slow job every 10 minutes. Each check cost ~20x a normal call (chart in the reply).
The fix is a sub-agent with a 1-hour cache. Save this as ~/.claude/agents/long-cache.md:
---
name: long-cache
description: General-purpose agent whose prompt cache lasts one hour instead of five minutes. Use it for agentic work that will sit idle more than five minutes between its own turns (waiting on long builds, test suites, visual checks, or its own subagents), especially once its context grows large, since every expired cache rewrites the whole context. For short or continuously active tasks use general-purpose: its cache writes cost less.
experimental:
cacheTtl: 1h
---
Work as a general-purpose agent on the task you're given.
Then ask Claude to use long-cache for anything that runs your slow commands.
Why not give every sub-agent a 1-hour cache? Its writes cost 60% more, and 53 of my 72 agents never waited 5 minutes. For them it's pure waste.
Unlike the main convo, sub-agents only keep their prompt cache for 5 minutes, even on a subscription. If one waits longer than that on a test t call pays for its whole context again. In that build it happened 200 times: 31% of the total cost.
The worst was one agent polling a slow job every 10 minutes. Each check cost ~20x a normal call (chart in the reply).
The fix is a sub-agent with a 1-hour cache. Save this as ~/.claude/agents/long-cache.md:
---
name: long-cache
description: General-purpose agent whose prompt cache lasts one hour instead of five minutes. Use it for agentic work that will sit idle more than five minutes between its own turns (waiting on long builds, test suites, visual checks, or its own subagents), especially once its context grows large, since every expired cache rewrites the whole context. For short or continuously active tasks use general-purpose: its cache writes cost less.
experimental:
cacheTtl: 1h
---
Work as a general-purpose agent on the task you're given.
Then ask Claude to use long-cache for anything that runs your slow commands.
Why not give every sub-agent a 1-hour cache? Its writes cost 60% more, and 53 of my 72 agents never waited 5 minutes. For them it's pure waste.
4
0
1
My Codex weekly usage after the release of Opus 5.5
Guess that's why I see less people asking @thsottiaux for resets this week?
Guess that's why I see less people asking @thsottiaux for resets this week?
1
0
1
New models edit files through the shell instead of the write tool all the time.
It's completely unreadable in Pi, no diff
In my agent transcripts: 30% of Opus 5.5's edits in Claude Code, ~20+% for GPT-6, GLM 5.3, DeepSeek v4 flash and Kimi K3 in Pi.
I built an extension that it. It parses the command with tree-sitter, nothing is executed, ~0.2 ms per edit command/script.
Your thoughts @mitsuhiko?
Here's the before/after, link below
👇
It's completely unreadable in Pi, no diff
In my agent transcripts: 30% of Opus 5.5's edits in Claude Code, ~20+% for GPT-6, GLM 5.3, DeepSeek v4 flash and Kimi K3 in Pi.
I built an extension that it. It parses the command with tree-sitter, nothing is executed, ~0.2 ms per edit command/script.
Your thoughts @mitsuhiko?
Here's the before/after, link below
👇
2
0
1
Ok I have to say Opus 5.5 blew me away
I asked it to build a full course to really understand LLMs down to the metal and be able to contribute to local AI on Mac (MLX, Metal)
It actually built and TRAINED an 80k params LLMs, I am now going through a full handcrafted course on model and it looks gorgeous and is super well built
I asked it to build a full course to really understand LLMs down to the metal and be able to contribute to local AI on Mac (MLX, Metal)
It actually built and TRAINED an 80k params LLMs, I am now going through a full handcrafted course on model and it looks gorgeous and is super well built
0
0
0
Why is Dario such a sinophobe?
Chinese labs have advanced open AI much more than anthropic afaik
plus Anthropic is probably using many chinese advances in their own models, but we will never know I guess
Chinese labs have advanced open AI much more than anthropic afaik
plus Anthropic is probably using many chinese advances in their own models, but we will never know I guess
0
0
1
Some providers on OpenRouter are really scummy
they use "discounts" so that all your traffic is routed to them by the default routing rule (cheapest first)
except the cache reads are 5 TIMES more expensive so you end up paying much more than you would with a seemingly more g: Morph, Silicon Flow
I recommend adding a guardrail in OpenRouter to blacklist them so that your requests are simply not routed to them. I also do that for providers that provide heavily quantized versions of models
they use "discounts" so that all your traffic is routed to them by the default routing rule (cheapest first)
except the cache reads are 5 TIMES more expensive so you end up paying much more than you would with a seemingly more g: Morph, Silicon Flow
I recommend adding a guardrail in OpenRouter to blacklist them so that your requests are simply not routed to them. I also do that for providers that provide heavily quantized versions of models
1
0
2
Here's Trueform, the skill and tools we put together.
Give your agent your own photos, or ask it to find references online. It guides the modeling process and helps catch shape errors across all angles.
Free and open source:
github.com/shidenkai0/tru…
Give your agent your own photos, or ask it to find references online. It guides the modeling process and helps catch shape errors across all angles.
Free and open source:
github.com/shidenkai0/tru…
0
0
0
I had Astra model this F-22 in Blender from photos only, then animate the takeoff. Afterburner and all.
We turned that workflow into a free skill and Blender tools to help your AI agent build 3D models that look right from every angle.
Get them below.
We turned that workflow into a free skill and Blender tools to help your AI agent build 3D models that look right from every angle.
Get them below.
1
0
1
This applies to scientific discoveries, AND to your own IP.
Don't forget to turn off "improve the model for everyone" in ChatGPT settings if you don't want your trade secrets to be embedded in the next model and become public knowledge.
x.com/ValerioCapraro…
Don't forget to turn off "improve the model for everyone" in ChatGPT settings if you don't want your trade secrets to be embedded in the next model and become public knowledge.
x.com/ValerioCapraro…
BREAKING: OpenAI might have stolen another major proof.
In a detailed Mastodon post, which I report in full in the comments, Andreas Thom presents several pieces of evidence suggesting that OpenAI may have trained Astra on conversations in which he and Gábor Kun were working on cture, one of the ten problems OpenAI later announced Astra had solved.
I know Andreas. We met several times early in our careers. He is an exceptional mathematician, a leading expert on sofic and hyperlinear groups, and one of the most respected scholars in the field. He has spent two decades working on this problem.
If his account is correct, this is not a minor dispute over attribution. It would mean that unpublished human work was absorbed into a model and then presented to the world as a breakthrough by the model itself.
And if the allegations raised by Levent Alpöge, Tristan Buckmaster, and now Andreas Thom are all substantiated, we are no longer looking at isolated incidents.
We may be looking at one of the greatest intellectual scandals in the history of science.
AI is not discovering new mathematics.
AI is stealing human discovery.
0
0
0
Let's go!
Btw if you don't have access to Astra yet, try using the apps or resetting your browser cache.
It still didn't show up for me on the web this morning, but was in the model list of both apps (mobile, Mac).
Btw if you don't have access to Astra yet, try using the apps or resetting your browser cache.
It still didn't show up for me on the web this morning, but was in the model list of both apps (mobile, Mac).
1
0
0
ChatGPT downtime update:
DNS records for chatgpt.com seem to be pointing to random IPs on the Internet, example
DNS records for chatgpt.com seem to be pointing to random IPs on the Internet, example
0
0
1
