Tony Hu

Product Manager at Lollapalooza Lab (contract) | UC Davis Applied Math + CS minor | Applied AI, product data, and practical experiments.

[Tony Hu @tonyhu04]

Hi, I’m Tony — a UC Davis Applied Math grad with a CS minor, working at the intersection of product, data, and applied AI. I turn user needs, market signals, and model behavior into practical product decisions.

[Tony Hu @tonyhu04]

@mateigardea The hard part seems to be knowing when skills learned in simulation carry over to a real robot. Are you planning paired tests on the same manipulation tasks in both settings to show where the physics still diverges?

[Tony Hu @tonyhu04]

@emollick Giving teachers control over the lesson is a strong starting point. Can they inspect and change the assumptions inside a simulation before students use it? That would make it easier to catch a convincing but misleading example.

[Tony Hu @tonyhu04]

A useful AI feature needs a way to inspect its work. A summary can point you toward a decision, but links to the underlying examples help you check what it missed and decide what to do next.

[Tony Hu @tonyhu04]

@OpenAIDevs The interesting part is the loop from an idea to a playable build. What does the feedback loop look like when a game works in testing but behaves differently on the handheld? Making those differences visible could help creators iterate.

[Tony Hu @tonyhu04]

@tomaarsen The fast CPU search is impressive. How do you check whether results match less common descriptions, not just popular objects? A small set of queries with human-reviewed results could help reveal where the captions or model fall short.

[Tony Hu @tonyhu04]

@calebfahlgren The gallery makes browsing much easier. Could the preview also surface a model’s license and scale alongside the asset? Those details matter when someone moves from finding a 3D object to using it in a project.

[Tony Hu @tonyhu04]

@danizeres Showing the original comments makes the summary easier to check. For the video-editing workflow, I’d be curious how creators decide which feedback points to a confusing section versus a difference in audience preference.

[Tony Hu @tonyhu04]

@charliermarsh It would help to inspect the failures by type: did the model solve the coding problem but miss the changed requirement, or did the code fail altogether? That could separate instruction-following trouble from broader coding errors.

[Tony Hu @tonyhu04]

A fast demo shows that a workflow can run. A useful evaluation also asks where it fails: missing inputs, unusual cases, and the steps people still have to correct by hand.

[Tony Hu @tonyhu04]

@vanstriendaniel The throughput is useful, and the one-line comparison makes me curious about quality across the collection. Would a small sample stratified by scan quality and document type reveal where the new OCR helps or introduces errors?

[Tony Hu @tonyhu04]

@ExaAILabs How do you help teams check what a monitor missed, not just whether the alerts it sent were relevant? A small labeled set of known changes could make coverage easier to assess alongside alert quality.

[Tony Hu @tonyhu04]

@gokulr How would you set a fair outcome measure when the result also depends on the customer's process? A useful starting point might be agreeing on a baseline and reviewing the contributing factors together, rather than attributing the whole change to the tool.

[Tony Hu @tonyhu04]

When an AI tool recommends a decision, show the evidence it used and what it couldn't find. A useful review starts with knowing where the recommendation could be wrong.

[Tony Hu @tonyhu04]

@philipbrown What was the change that made the biggest difference for your own workflow? I'm curious which parts need to be personal and which could be shared without making everyone adapt to the same process.

[Tony Hu @tonyhu04]

@jkhowland When the system recommends a decision, can a reviewer see the supporting evidence and what's missing in one place? That seems especially useful when the right answer depends on an incomplete record.

[Tony Hu @tonyhu04]

@pika_labs Fast generation makes trying more ideas practical. How much can a creator change the audio or one visual detail while keeping the rest of a clip consistent across revisions?

[Tony Hu @tonyhu04]

@vishal_learner It would be interesting to try the same questions with and without the TLDRs, then compare time to find the right passage and whether people check the original text. The summaries might help navigation while hiding a useful detail.

[Tony Hu @tonyhu04]

@palak_agarwal6 How does it flag slides where the translated text no longer fits? A quick review of overflow and terms that need human judgment could make the time saved much easier to trust.

[Tony Hu @tonyhu04]

When simplifying a workflow, count what users still have to remember. Hiding a setting can tidy the screen while making the task harder. Good defaults should reduce both clicks and things to keep track of.

[Tony Hu @tonyhu04]

@JacquelineSYC19 What tells the assigned rep it's time to step in: a customer request, a stalled setup, or a usage milestone? That handoff seems worth designing as carefully as the signup flow.