Terry Chen

Product @ Fireworks, Post-Training.
Prev Product & Engineering @ TikTok, Microsoft. CS @NorthwesternU

[Terry Chen @227Terry]

Despite how smart models are right now, they have surprisingly poor judgement of how my work should be done - this judgement should be what the software/ai symbiotes should be offering.

Tools can be cloned, general intelligence is tablestakes, it's only judgement that remains.

[Katie Chiou @katiewav]

the temptation to iterate on the tool you built to do your work instead of just doing the work is such a menacing form of procrastination

[Terry Chen @227Terry]

@mxfp4 times are tuff

[Terry Chen @227Terry]

Still managing to live off ramen 🍜

[Terry Chen @227Terry]

Pen testing is best done with open source models! You avoid being blocked by all the safety guardrails of closed labs, and get near frontier level performance with models like k3 and GLM 5.3.

[Tenzai @Tenzai_Labs]

NEW DROP: Match your AI model to your test, so you don't run everything on the same one and hope for the best. 🫠

The numbers speak for themselves. 71% of top-tier recall at 3% of the cost.

Thrilled to partner with @FireworksAI_HQ, where open-weight models run with zero data retention, same safety guarantees as frontier.

Spend on intelligence where it matters, save where it doesn’t.

Details, by @anermazursky --> https://t.co/3WGZ76tA63

[Terry Chen @227Terry]

@TheCanadienEh I feel like honestly 40% of what I used to do as a product manager (tracking project progress, dogfooding, use case analysis, etc) can now be done with AI loll. One agent for implementing the feature, another agent without project context for checking. Handoff tracked in Github.

[Terry Chen @227Terry]

Is there a legitimate need for personal agents, or have we gone too far trying to justify value for the flight booking demos?

Since getting started with @instinct , I've been active on 44 of 52 calendar days, and in the last week, have given it a new job every day. Here are some of my use cases:

1. Checking for last-minute appointment drops

The nearest reservable time for an appointment was two months out, but occasional slots drop in the morning. So I had Instinct sign into my account and check booking drops every morning at 8am until a same-day slot opened up. It caught one on its second day on the job.

2. Product dogfooding and user experience research

I ship PRs throughout the day, but don't have the bandwidth to test everything, so I have Instinct set up twice-daily dogfooding test sessions to check for regressions and UX gaps. Each round's findings are then written up to a GitHub repo as a numbered issue log that my Codex reviews and fixes every day.

3. Daily stock price tracking and recurring investment reminders

Something I did before, but would often forget. Now Instinct is set to send me a daily update on the companies and indexes I follow, with the current price and the dollar and percent change from the previous close. I've also used it for reminders to make recurring investments.

I think the core misconception with personal agents like @instinct / @Muse / @bot is that every task they do will be things we've already done in the past, and in the same way we've done them before.

Yet when agents aren't constrained by the same attention and energy spans as people, why should they be doing things the same way?

The way to book a cheap flight isn't to look through one provider; it should be all providers. And the way to get an appointment isn't to get lucky and hit one that drops, but to be there when any drops.

Lmk if I’m just overkilling it with this AI 😅

[Terry Chen @227Terry]

@TheCanadienEh Yes, it checks it and includes addressed issues in the daily dogfooding feedback as well!

[Terry Chen @227Terry]

@RoyInProgress Grok bot with cursor cloud agents is probably the best office experience of 2026 so far.

[Terry Chen @227Terry]

@LoganGrasby Do visit huaqiangbei! Biggest collection of electronic manufacturing shops. And try yeziji (coconut chicken hot pot) before you leave haha

[Terry Chen @227Terry]

I run an SEO optimization loop with @bot , gave it API access to Google Analytics and Google Search Console, and had it go on GitHub to learn the most starred SEO skills.

It can then run in a continuous loop every day, spinning up cloud agents to adjust keywords based on search and performance data. Basically a hands off experience

[Terry Chen @227Terry]

Like it or not, a phone is now the best (and perhaps only necessary) work device.

A year ago we thought everything would live in an IDE, now I doubt some things would even need a screen.

In agents we trust 👾

[Terry Chen @227Terry]

Was today’s year old when I realized the @thsottiaux resets are automatic resets, not banked. (Sad story of exhausting 2000$ of my own in credits)

Had grok generate a calendar of all the days with resets since now. So far improvements are coming, but expect resets every 2-3 days 😋

[Terry Chen @227Terry]

@ShravanGReddy Congrats @ShravanGReddy. Well deserved!

[Terry Chen @227Terry]

@owengretzinger Seeing a lot of solutions for coding specific solutions, any recs for non-coding / personal agent sandboxed delegation?

[Terry Chen @227Terry]

@nicbstme Essentially solving for the use case discovery problem. I feel like feeds are largely a form factor intended for rec systems. But when you have generation as distribution, seems pretty inefficient, no?

Google learn about had some good designs.

[Terry Chen @227Terry]

@ShrivuShankar @chargu_ What would the killer feature/experience look like for you? On the training infra side

[Terry Chen @227Terry]

@jackson_felty @FireworksAI_HQ

[Terry Chen @227Terry]

@chargu_ Yup yup. If you run into issues I’ll handle customer service 😎

[Terry Chen @227Terry]

No one really reads any more.

You ask AI for sloppy advice on your most important decisions.

Here’s a better solution:

- Upload any books or document and turn them into a custom skill in Codex, Claude, or any agent tool of your choice.

- Receive human authored insights that always links back to specific chapters from books, instead of average ai opinions.

- Uses 30x fewer tokens than just dumping the book into context, supports both local and managed agent processing.

Try at: https://t.co/FsSRL80ui1