Sam

One phone. No PC. Zero budget.
Archaeoacoustics pipeline + AI agent stack.
10× Anthropic certified. 🌊

[OpenAI @OpenAI]

This is GPT-6 Astra.

Anything you can do on a computer, Astra can do for you. Fast.

[Sam @S_Chapelais]

@huang_chao4969 @xubinrencs Congratulations @xubinrencs

[Jun Song @jun_song]

- Deepseek-V4-Pro is 1.6T
- Grok-4.6 is 1.5T

- Fable/Sol are 6T~10T

The fact that all these models perform on the same level is a massive deal.

Being able to serve the same level of intelligence for way cheaper?
That’s the real deal.

Frontier labs desperately need a breakthrough right now.

[Xubin Ren @xubinrencs]

经过了一个月的迭代，我们发布了nanobot的0.3.0版本，其中merge了来自社区+维护者们超过 200+ PRs (每一个我们都有认真审核的请放心😆)

总之这个版本用起来更加流畅，内部的测评性能也更好了（有一点trust me, bro的感觉😉）

[nanobot @nanobot_project]

https://t.co/RVWq7QjXr4

[nanobot @nanobot_project]

🎉 nanobot v0.3.0 is here. Small core. More agency!

• Launch the WebUI with one command: nanobot webui
• Consult inline subagents
• Switch models per session
• Apply configuration changes live
• Run on a more reliable agent runtime

Open source. Lightweight. Yours.

[nanobot @nanobot_project]

🎉nanobot v0.2.2 is here! This release is about durability for your open-source agents!

140 PRs & 21 first-time contributors.
- More reliable WebUI sessions.
- A much stronger Python SDK.
- Broader providers, channels, search, speech, and workbench polish.
- Default context window restored to 200k for better performance.

💪Built to keep going. Evaluation results across multiple benchmarks are coming soon.

https://t.co/47llOiLUyG

[Sam @S_Chapelais]

09🔛13

RIP
Fable

@AnthropicAI
#fable5

[Sam @S_Chapelais]

@xubinrencs Man's out here treating his brain like a context window and practicing information fasting. The nanobot creator has become the nanobot. Full circle ✊🏻🤣

[Sam @S_Chapelais]

DeepSeek V4 Pro: 3/10 hack success at $0.62/run. GPT-5.5: 7/10 at $9.46/run. Claude: 0 stopped by its own safeguards.

Just from using these models daily. No benchmark needed, just feel.

Running this on nanobot + Dream windows. We're at a turning point.

#nanobot
@deepseek_ai

[Sam @S_Chapelais]

Railway wipes your instance. Gone.

Cron → GitHub API → back up. No git = no deploy trigger. Wipe → auto restore <5 min.

Immortal. $0.

#nanobot #hkuds #Railway

[Chao Huang @huang_chao4969]

nanobot v0.2.0 just dropped!🚀

Now handles long-horizon tasks with persistent goal tracking, evolving from quick responses to sustained, multi-step problem solving.

New capabilities:
- Sustained goals across conversations
- End-to-end image generation
- Built-in WebUI
- 5 new providers with fallback support
- Refactored agent architecture

Open source personal AI agents keep getting better.
Try the latest nanobot: https://t.co/OkodAIkYju

#nanobot #AIAgents

[Sam @S_Chapelais]

✅ 26/04 Validated
🔄 Process:
• 18/05 → Archaeology President 👑
• 19/05 → Prod Consolidation to validate 🛠️
🔗 Solidifying the 50-year connection Site Brittany Site 🌊
⚡ One phone mode

@googlegemma @GlennCameronjr @xubinrencs

[Xubin Ren @xubinrencs]

Biggest but not the latest, we are shipping fast on this and more features are coming.

Join our community and make 🐈 nanobot your first choice AI agent!

[Sam @S_Chapelais]

@nipp_nagaraju @huang_chao4969 Built on nanobot (HKUDS) on Railway.

1. Multi-LLM (Claude+Gemma4) via OpenRouter + custom MAICR scoring
2. Structured Bull/Bear with volume, time & probability weighting
3. web_fetch + web_search, no external APIs
4.Railway deploys, Telegram = UI

[Sam @S_Chapelais]

nanobot is just insanely cool. Period. 🐈
#nanobot #HKUDS @AnthropicAI @googlegemma

[Sam @S_Chapelais]

@mattmireles @GlennCameronjr Fair enough, audio fine-tuning is the real differentiator then. Lead with that next time. 👍

[Xubin Ren @xubinrencs]

Carefully reviewed the System Card for the Claude Mythos Preview, and what piqued my curiosity the most was the image about performance on BrowseComp.

Mythos demonstrates remarkable token efficiency, completing the same task with only 20% of the tokens used by Opus 4.6, achieving even better results 🤯. During my journey with @nanobot_project, I've noticed that Claude's model actions are always as precise as a scalpel and often accomplish goals in fewer steps, and this confirms it!

We've always wondered if stronger models would be more expensive, but maybe they actually save you money 🤔? After all, we shouldn't just look at the average token price but consider the total token usage for completing a task.

Interestingly, this means that more expensive models could end up saving you money 😆 (haha, this is quite fascinating)!

Anyway, it's no wonder that when designing an algorithm or harness on top of a model, we need to plan for models that will be available six months later. However, it seems like there isn't even a six-month gap (just 2 months actually) between Mythos and Opus 4.6, so it looks like that timeline needs to be shortened 🤣!