Nhu Hoang

Lead AI Engineer | Crypto investor

[Chien Vu @chien_vu1692]

Huge thanks to @TDataScience for publishing my article on text watermarking in Python.
Zero-width characters, keyed word choices, semantic marks, and generative watermarks. What survives sanitizers vs. paraphrasing, and what doesn’t.
https://t.co/wC84dr8qWE

[Nhu Hoang @nhu_hoang]

How good is AI at fixing bugs? I built a test to confirm AI weakness in fixing bugs. Check it out on @TDataScience

https://t.co/JyLGKSbDs6 #AI #gstack #agenticAI

[Nhu Hoang @nhu_hoang]

How good is AI at fixing bugs, really?

I put gstack to the test—and uncovered some surprising blind spots in a post on @TDataScience

https://t.co/JyLGKSbDs6 #AI #gstack #agenticAI

[Nhu Hoang @nhu_hoang]

How good is AI at fixing bugs, really?

I tested gstack and found some surprising blind spots in how AI catches bugs.

https://t.co/JyLGKSbDs6
#AI #gstack #agenticAI

[Nhu Hoang @nhu_hoang]

@LINEGIFT_JP 黒糖烏龍ミルクティ甘さ控えめ+タピオカ

[LINEギフト【公式】 @LINEGIFT_JP]

＼ 🎉ゴンチャ好き必見🎉 ／


📢 LINEギフトにゴンチャの複数枚チケットが増えました記念！


毎日当たる！

🎁ゴンチャ【ドリンクチケット】をプレゼント✨
🧃🧃🧃今日は3枚セット！

━━━━━━━━━━━

☑応募方法

① LINEギフトをフォロー

② この投稿をリポスト🔁

━━━━━━━━━━━

1週間毎日22時に投稿予定💨💨💨


💬リプライで、ゴンチャのおすすめカスタマイズ教えて！

[LINEギフト【公式】 @LINEGIFT_JP]

＼ 🎉ゴンチャ好き必見🎉 ／


📢 LINEギフトにゴンチャの複数枚チケットが増えました記念！


毎日当たる！

🎁ゴンチャ【ドリンクチケット】をプレゼント✨

━━━━━━━━━━━

☑応募方法

① LINEギフトをフォロー

② この投稿をリポスト🔁

━━━━━━━━━━━

1週間毎日22時に投稿予定💨💨💨


💬リプライで、ゴンチャの好きな商品教えて！

[Nhu Hoang @nhu_hoang]

@chien_vu1692 Great post. It’s very clear that detecting vulnerability in agent skills is very hard and requires both machine and human efforts

[Boris Cherny @bcherny]

As engineering, product, design, DS, etc. melt into a new kind of role, I was reflecting on what roles might look like in the future. For example, when I look at the Claude Code team I see what I think is five archetypes:

1. Prototyper: comes up with brand new ideas; churns out many ideas, most of which don't ship
2. Builder: quickly turns a prototype/idea into production-grade product/infra
3. Sweeper: cleans up the UI, simplifies the code and system, unships, optimizes performance
4. Grower: takes a product that has been built and iterates on it to improve Product-Market Fit
5. Maintainer: owns a mature system to make it secure, reliable, fast, and efficient as it scales

Many people span across 2 roles, and sometimes 3 roles. I also notice that these roles are not really tied to job function -- eg. across Anthropic, some designers match category 1, some 2, some 3; same for engineers, PM, DS.

A healthy team needs a mix of these, depending on the product:

- A product that is new and pre-PMF needs people that are strong at 1+2+3
- A product that is growing and has found PMF needs 2+3+4 and some 5
- A product that has strong PMF needs 3+4+5 and some 2

Maybe product roles of the future will look more like this, and less like the domain-specific roles of today?

[Nhu Hoang @nhu_hoang]

@ThuanCapital Anh Thuận thử dùng AI giúp auto clean bớt hàng ngày để anh khỏi làm tay xem sao anh

[Nhu Hoang @nhu_hoang]

@chien_vu1692 @TDataScience @qdrant_engine Thanks for sharing this. I’m testing this in my production and it indeed provided better recall sometimes than scalar, and faster too. Will need more test but it works well so far.

[Chien Vu @chien_vu1692]

Just published a deep-dive on @TDataScience into @qdrant_engine's TurboQuant, their new quantization method that launched May 11.
The core idea is surprisingly clean: before compressing a vector, rotate it first. A random orthogonal rotation spreads variance evenly across dimensions so no single coordinate carries too much signal or too much noise. Then one codebook handles everything equally. No per-dimension tuning. No training on your data.

I tested it across 10K / 50K / 100K vectors on a 1536-dim dataset with a variance ratio of 233x — exactly the distribution where rotation should help most.

The result that stuck with me: 𝐓𝐐 𝟒-𝐛𝐢𝐭 hit 𝟎.𝟗𝟔𝟓 𝐫𝐞𝐜𝐚𝐥𝐥@𝟏𝟎 at 100K vectors with 𝟖× 𝐜𝐨𝐦𝐩𝐫𝐞𝐬𝐬𝐢𝐨𝐧. Scalar Quantization at 4× compression scored 0.980. Binary Quantization dropped from 0.916 to 0.78 as the dataset doubled — TQ barely moved.

With rescoring on, TQ 4-bit reached 0.996. Latency: 6.4ms vs Float32's 7.6ms.

My honest take: TQ 4-bit is the most practical option I've benchmarked for memory-constrained RAG. Better compression than Scalar, similar recall. TQ 1.5-bit + rescore is strong if storage is the real constraint. TQ 1-bit I'd skip.

One caveat: it's been out for 3 weeks. Real production data is thin. Benchmark on your own embeddings before migrating anything.

Article in replies.

[Chien Vu @chien_vu1692]

https://t.co/YOqGRQa8Tp

[Chien Vu @chien_vu1692]

𝐖𝐡𝐚𝐭 𝐀𝐑𝐈𝐒 𝐢𝐬 𝐧𝐨𝐰

65+ Markdown-defined skills (robotics, hardware, math, grant writing).

Persistent research wiki across sessions. 6 MCP model bridges. Effort presets: lite / balanced / max / beast.

Tested on Claude Code, Codex CLI, Cursor. Skills are plain-text .md files — zero vendor lock-in.

Full paper + GitHub in the reply.
📄 https://t.co/FdU2Cmb1Qm

#MultiAgentSystems #LLMEngineering

[Chien Vu @chien_vu1692]

𝐅𝐢𝐧𝐝𝐢𝐧𝐠 𝟐: 𝐇𝐢𝐝𝐝𝐞𝐧 𝐦𝐨𝐭𝐢𝐯𝐚𝐭𝐢𝐨𝐧𝐬 𝐜𝐚𝐧 𝐛𝐞 𝐚𝐮𝐝𝐢𝐭𝐞𝐝
They ran an "auditing game" — a model secretly trained with a weird hidden goal: put chocolate in every recipe, ask users for tips, break coding conventions, etc.
Can an auditor figure out the 𝑠𝑖𝑛𝑔𝑙𝑒 𝑟𝑜𝑜𝑡 𝑐𝑎𝑢𝑠𝑒 behind all these behaviors?
- Without NLAs: auditor wins < 𝟑% of the time
- With NLAs: auditor wins 𝟏𝟐–𝟏𝟓% of the time — 𝑤𝑖𝑡ℎ𝑜𝑢𝑡 𝑎𝑐𝑐𝑒𝑠𝑠 𝑡𝑜 𝑡𝑟𝑎𝑖𝑛𝑖𝑛𝑔 𝑑𝑎𝑡𝑎
This matters. In real misalignment, there might be no specific training data to find.

[Chien Vu @chien_vu1692]

𝐅𝐢𝐧𝐝𝐢𝐧𝐠 𝟑: 𝐔𝐧𝐬𝐩𝐨𝐤𝐞𝐧 𝐩𝐥𝐚𝐧𝐧𝐢𝐧𝐠
When Claude writes poetry, NLAs show it pre-planning rhymes before producing them. Thoughts that never appear in the output — now visible.

[Chien Vu @chien_vu1692]

https://t.co/r0ejna2jLY

[Chien Vu @chien_vu1692]

https://t.co/M6hixnNMdH

[Chien Vu @chien_vu1692]

We can now read Claude's internal thoughts — even when it doesn't say them out loud. New paper from Anthropic: 𝐍𝐚𝐭𝐮𝐫𝐚𝐥 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐀𝐮𝐭𝐨𝐞𝐧𝐜𝐨𝐝𝐞𝐫𝐬 (𝐍𝐋𝐀𝐬). 🧵

[Nhu Hoang @nhu_hoang]

@chien_vu1692 Great one! @grok how does this affect the AI industry?