Chau Tran

1.1K posts

Chau Tran

Chau Tran

@mr_cheu

Building AI @glean. Past: Machine Translation at FAIR, ML @Quora

New York, USA Katılım Nisan 2009
943 Takip Edilen1K Takipçiler
Chau Tran
Chau Tran@mr_cheu·
What's the best opensource code mode tool execution that's lightweight, support both MCP and non-MCP tools? Don't want dependencies on any agents/harness sdk
English
1
0
1
176
Chau Tran
Chau Tran@mr_cheu·
@giffmana Maybe crashes are penalized more than silent failures in post-training?
English
0
0
0
115
Lucas Beyer (bl16)
Lucas Beyer (bl16)@giffmana·
PLEASE Anthro and OpenAI, make sure this paper is in the training data. He'll put it in the context! I'm tired of the code you generate "failing silently" instead of loudly so i can investigate! Don't be robust to document.querySelector() not finding the thing we *need* to exist. Just fail loudly so we notice and fix, instead of going down a long hunt for "why nothing happens". (and Google too? Haven't tried gemini-cli yet)
Dominik Tornow@DominikTornow

assert() in production, crash on violation An assertion violation means the component already failed, just hasn't crashed yet Crash the component and let the rest of the system compensate A crash is always a possible failure mode Take advantage

English
26
25
576
104.5K
Han
Han@HanchungLee·
anthropic api is taking over chat completion. its natively agentic, supports tool use, and adopted by almost all chinese labs (the only worthy oss players). at the same time, oai left chat completion to die in favor of response api.
Skyler Miao@SkylerMiao7

Hey guys, thank you all for your love and passion for M2. Lot asking why we recommend Anthropic API. I think need to explain a little bit. M2 is a agentic thinking model, it do interleaved thinking like sonnet 4.5, which means every response will contain its thought content. Its very important for M2 to keep the chain of thought. So we must make sure the history thought passed back to the model. Anthropic API support it for sure, as sonnet needs it as well. OpenAI only support it in their new Response API, no support for in ChatCompletion. That's why GPT5 has best performance only under Response API.

English
5
6
186
28.6K
Chau Tran
Chau Tran@mr_cheu·
Claude Code/Codex should evolve to become a full terminal replacement, accepting both bash scripts and natural language instructions
English
0
0
2
146
alex zhang
alex zhang@a1zhang·
What if scaling the context windows of frontier LLMs is much easier than it sounds? We’re excited to share our work on Recursive Language Models (RLMs). A new inference strategy where LLMs can decompose and recursively interact with input prompts of seemingly unbounded length, as a REPL environment. On the OOLONG benchmark, RLMs with GPT-5-mini outperforms GPT-5 by over 110% gains (more than double!) on 132k-token sequences and is cheaper to query on average. On the BrowseComp-Plus benchmark, RLMs with GPT-5 can take in 10M+ tokens as their “prompt” and answer highly compositional queries without degradation and even better than explicit indexing/retrieval. We link our blogpost, (still very early!) experiments, and discussion below.
alex zhang tweet media
English
135
378
2.7K
947.6K
Chau Tran
Chau Tran@mr_cheu·
gpt-5-pro API wen?
Indonesia
0
0
2
278
Jo Kristian Bergum
Jo Kristian Bergum@jobergum·
Wonder what happened with the @LinkedIn recsys algo that now prioritizes 2-3w old posts that I have seen before. Congratulations!
English
7
1
32
3K
Chau Tran
Chau Tran@mr_cheu·
@headinthebox Probably because this model was post trained with tools in that exact schema? Another generalized version should come later
English
0
0
0
98
Erik Meijer
Erik Meijer@headinthebox·
OpenAI Deep Research MCP support: ... Your MCP server must implement two tools to work with deep research - search and fetch ... [0] This feels counter to the spirit of MCP, and I think using regular function calling is more appropriate.
English
5
3
20
4K
Chau Tran
Chau Tran@mr_cheu·
@walden_yan It depends on whether you’re building a vertical agent or horizontal agent though. Broad horizontal agents like ChatGPT or Glean benefits more from calling specialized subagents
English
0
0
0
200
Walden
Walden@walden_yan·
I see a lot of people make the same mistakes building agents. So we shared a few of the principles we use cognition.ai/blog/dont-buil…
English
61
126
1.1K
248K
Chau Tran
Chau Tran@mr_cheu·
We need an LLM API that can iteratively call tools, see tool outputs, call tools again,… until finally streaming back the response
English
1
0
0
266
Chau Tran
Chau Tran@mr_cheu·
4/n What are the best memory management techniques for long-running agents, and what we need from future LLMs
English
0
0
0
132
Chau Tran
Chau Tran@mr_cheu·
3/n When to use finetuning vs dynamic prompting for training agents that scale and personalize
English
1
0
0
159
Chau Tran
Chau Tran@mr_cheu·
🚀 Excited to speak at @aiDotEngineer this Wednesday @ 2pm! Main topics: 1/n How to build enterprise-aware agents that are not just smart but also adaptable and aligned with your company’s processes
English
3
0
7
303
Chau Tran
Chau Tran@mr_cheu·
Just look at the system prompts from all the frontier labs if you don't believe me
English
0
0
1
107
Chau Tran
Chau Tran@mr_cheu·
Prompting is like writing software. Finetuning is like building customized hardware. The first is more flexible, you can adapt to the requirements very quickly The second is more optimized, but more costly, so you can only do it when the requirements don't change too quickly
Aaron Levie@levie

The more time you spend with AI the more you realize prompt engineering isn’t going away any time soon. For most knowledge work, there’s a very wide variance of what you can get out of AI by better understanding how you prompt it. This actually is a 21st century skill.

English
1
0
1
384

Keşfet

@giffmana @ScottWu46 @antigravity @HanchungLee @a1zhang @lateinteraction @trq212 @raveeshbhalla