Chau Tran

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

[Chau Tran @mr_cheu]

Is anyone else seeing the new Claude models (Sonnet 5, Opus 4.7, 4.8) tokenize into ~60% more tokens than GPT models for the same input prompt?

So per-token-pricing for Claude has to be multiplied by ~1.6 before comparing with GPT

This is just counting input tokens, not output

[Chau Tran @mr_cheu]

@mweinbach GLM 5.2 inference can be much cheaper (than pay-per-token price) when done on dedicated compute

[Chau Tran @mr_cheu]

@jietang Image understanding

[Chau Tran @mr_cheu]

Ok so they CAN look at ZDR requests

[Chau Tran @mr_cheu]

how do they differentiate between normal requests and distillation requests, assuming they can't "look" at the requests with ZDR enabled?

[Chubby♨️ @kimmonismus]

Anthropic claims: Alibaba continues to distill Claude on a large scale to train Qwen. Via Bloomberg

Anthropic is accusing Alibaba-linked operators of running a massive campaign to illicitly access Claude through nearly 25,000 fraudulent accounts.

According to Bloomberg, Anthropic claims the campaign generated 28.8 million Claude exchanges between April and June, targeting capabilities like software engineering and agentic reasoning.

The company says this is part of a broader pattern of “adversarial distillation,” where Chinese labs allegedly harvest outputs from US frontier models to train rival systems at a fraction of the cost.

Lets see how good Qwen 3.8 will be, probably FABLEous good.

[Chau Tran @mr_cheu]

The new prompting meta of Jun 2026:

"We got this bug report [..]. Send a test request to dev server to repro. Retrieve production logs to diagnose. Read our code base and implement a fix. Deploy your fix to dev server. Send a test request again and iterate until fixed”

[Chau Tran @mr_cheu]

With all due respect, Slack bot is not a newly invented concept

[Tony Gentilcore @tonygentilcore]

https://t.co/4z4aAS9VKl

[Chau Tran @mr_cheu]

@tuhinone Very nice speed. Is it true that context window is just 131k tokens though?

[Chau Tran @mr_cheu]

It is so refreshing to see full unsummarized thinking trace again. Makes it much easier to debug context/prompting issues #GLM52

[Chau Tran @mr_cheu]

GLM-5.2 is apparently less censored than some other Chinese models. It doesn't reject questions about Tiananmen Square 1989

[Chau Tran @mr_cheu]

If an LLM can recursively self-improve inside a frontier lab, can it do the same outside one? If so, releasing an RSI-capable model would effectively destroy the lab’s own moat

[Chau Tran @mr_cheu]

@suchenzang sadly there's no verifiable reward for writing quality. even high quality human preferences feedback is hard to get?

[Chau Tran @mr_cheu]

Are we just going to accept that Claude Code now just sometimes fail to do basic things like making a file edit? Started happening with Opus 4.7

[Chau Tran @mr_cheu]

@hwchase17 @nikunj Why can't LLM be the adapter between any arbitrary memory format?
Tell System A: Give me everything you know about me in a markdown file
Tell System B: [upload the markdown file] This markdown file contains everything System A knows about me, import it

[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

[Chau Tran @mr_cheu]

https://t.co/Xp0MpQr3cI

[Chau Tran @mr_cheu]

@giffmana Maybe crashes are penalized more than silent failures in post-training?

[Chau Tran @mr_cheu]

@ScottWu46 @antigravity This is embarrassing.

[Chau Tran @mr_cheu]

@HanchungLee what's wrong with responses api?