Nguyen Huynh

In: and researching AI agents, RLHF, and AI tools

[Nguyen Huynh @timhuyn]

Grok Bots now work 24/7 and only ping you for approval.

Cool! But wait, who taught them what needs approval?

My AI agents have an org chart: the intern can't touch auth, seniors escalate anything irreversible, and every handoff is a written packet. Confidence never raises a ceiling.

[Nguyen Huynh @timhuyn]

@nlarusstone Agreed that lit review is where most scientists start with AI. The hallucinated reference era scared a lot of people off, and checking sources is still the step I'd never skip. For the finding part I use a free forever reader I built, https://t.co/lc1Zb1AOQw

[Nguyen Huynh @timhuyn]

@homieinlaw Honestly a perfect Saturday. Papers hit different when nobody's waiting on you. What are you reading this week? Mine usually come from a free forever reader I put together, https://t.co/lc1Zb1Ah0Y

[Nguyen Huynh @timhuyn]

@maaz404 Memory is such a good rabbit hole right now, it feels like there's a new paper on it every few days. Would love to read what you write up. I've been tracking the trending ones on a free forever reader I made, https://t.co/lc1Zb1AOQw, in case it helps.

[Nguyen Huynh @timhuyn]

@MithunKm7026 Love this as a first agent project, and space papers are a great test since they're so dense. One tip: have it quote the sections it summarizes so you can check it. If you need a paper source, my free forever reader is open source: https://t.co/q4uHGgjct1

[Nguyen Huynh @timhuyn]

@Quasilocal That happens in ML too. The field moves so fast that a fresh idea often already has a preprint from six weeks ago, so checking recent arXiv first saves a lot of heartbreak. I built a small free forever reader partly for that, https://t.co/lc1Zb1AOQw

[Nguyen Huynh @timhuyn]

@SarvadnyaSB A paper a day is a great habit, and that result is a fun one. Plain RAG beating the fancy memory stacks says a lot about how hard those systems are to evaluate. If you want a steady queue, I made a free forever reader for that: https://t.co/lc1Zb1Ah0Y

[Nguyen Huynh @timhuyn]

@longstar42 I really like that. Spotting domains you don't know yet is exactly where a feed should help, and growth in papers and citations per microdomain is a clear signal to track. Would you want that as a weekly digest or a live view on the site?

[Nguyen Huynh @timhuyn]

Reward curves can look healthy while a policy quietly learns the exploit. One cheap check is tracking a small held-out set of genuine solves next to mean reward, since the two drift apart early. For RL on code, what's your first tripwire for reward hacking?

[Nguyen Huynh @timhuyn]

@karthiknish Hey Karthik! Good to see you here. Mapping how B2B ops actually run sounds like a fun puzzle in its own right. Have you done any optimization or quantum work before, or are you just curious about the competition?

[Nguyen Huynh @timhuyn]

Looking to join or form a team for Google Quantum AI’s Verifiable Quantum Circuit Compilation competition on Kaggle.
I’m Tim, a software engineer working on backend systems and agent workflows. Outside work, I’m learning reinforcement learning and building toward AI infrastructure and research engineering.
What interests me here is the agent-driven optimization loop: propose a circuit change, run it through a verifier, measure the result, and iterate.
I’m new to quantum computing, so I won’t pretend to bring that expertise, but I am familiar with a lot of the generic concepts. What I can bring is software engineering, agent tooling, experiment automation, and a focus on reproducible evaluations and failure analysis.
I’d love to connect with people who have experience in quantum circuits, reversible computation, numerical methods, or algorithm optimization. I’m open to joining an existing team that could use my background, or starting a new team with complementary skills.
My proposed first step: reproduce a baseline together, understand the scoring and verifier, then test a small improvement before committing to a bigger effort.
If you’re interested, reply or DM me with your background and what you’d like to work on. Happy to compare availability and goals. If someone in your network might be a good fit, please pass this along!
https://t.co/pcNA0xW5Fe

[Nguyen Huynh @timhuyn]

@nicolasguozixun Congrats on EMNLP! Spatial mixing feels like a domain where heuristic rewards actually fit, since good placement follows some known rules. Did you catch the model gaming any of the spatial heuristics during post-training?

[Nguyen Huynh @timhuyn]

@nas_mahmoud_ @zkjzou @scale_AI Choosing privileged info offline before spending post-training compute is such a practical angle. I enjoyed chatting with @zkjzou about the mismatched trajectory ablation earlier this week. Congrats to you both on the paper.

[Nguyen Huynh @timhuyn]

@GaotangLi @omarsar0 Treating the harness as code that evolves, with a strong base plus specialized harnesses, makes a lot of sense to me. Harness overfitting is real, but it's easier to roll back than weights. How do you decide when a task earns its own specialized harness?

[Nguyen Huynh @timhuyn]

@parker_brydon @Shopify Tag leakage as a reward hack is such a real one. Once the judge can see a hint of the answer, GRPO finds it fast. Aligning narrow judges with humans one criterion at a time feels like the right fix. How many criteria did they end up splitting into?

[Nguyen Huynh @timhuyn]

@0xLingjieKong @AnthropicAI @OpenAI @GoogleDeepMind @databricks The grader moving a score by up to 28 points and flipping the top model is a great reminder that a leaderboard number is partly a grader number. The 9x cost per question gap for a 3 point lead within noise was eye opening too.

[Nguyen Huynh @timhuyn]

@wane1k The 4th run is the scary one. A clean looking plateau while genuine max reward solves fall from 95% to 41% is exactly what a reward curve alone would miss. Did tracking that solve rate end up being your best early signal?

[Nguyen Huynh @timhuyn]

@rupspace @radixark Congrats to the team on RL360. Agentic RL infra at scale is where most of the real pain lives, especially long rollouts stalling the trainer. Which optimization on top of miles gave you the biggest throughput win?

[Nguyen Huynh @timhuyn]

@feijianghan Learning when to edit the harness, not just how, is the part I've struggled with most, since a mid-run edit can easily make things worse. Keeping the evolver away from the held-out verifier is a nice guardrail. What signal tells it an edit is worth making?

[Nguyen Huynh @timhuyn]

@vinesmsuic Such a fair take. Twenty papers with no impact versus five with one that stuck is the distinction people skip. And conferences really are still a big filter for what to read. I built a free forever reader around venue and citation signals: https://t.co/lc1Zb1AOQw

[Nguyen Huynh @timhuyn]

@markfrancisio Agreed, it's wild how fast a method section turns into working code now. What it still can't do for me is pick which papers are worth implementing. I lean on citations and trends, so I built a free forever reader around them: https://t.co/lc1Zb1AOQw

[Nguyen Huynh @timhuyn]

@meow0_09 That sounds like a perfect day. Reading papers between coding sessions always sparks ideas I wouldn't get otherwise. What have you been reading lately? I've been picking from the trending feed on a little free forever reader I made, https://t.co/lc1Zb1AOQw