Nathan Zylbersztejn

Engineering

[Nathan Zylbersztejn @nathan_zyl]

@nikitabier Why don’t you tell us more about your time rioting with young over educated communists in the 2010s ?
Btw, French protesters are high school students

[Nathan Zylbersztejn @nathan_zyl]

@levelsio There was never a moat, there was a gate (fewer coders, higher costs ). AI is to software what internet was to music. You don’t need a label to get started, but you’ll get one after initial distribution is secured. Is the music business dead?

[Nathan Zylbersztejn @nathan_zyl]

@mattpocockuk Back in 2018 most teams were agile/scrum, implementation decisions were refined and consensual, and each pr was reviewed. Where were the heroes during grooming and reviews ?

[Nathan Zylbersztejn @nathan_zyl]

@bcherny @lucas_builds Also, Claude reviews are ridiculously expansive (15-25$) while dedicated vendors charge 1-10$ and do a pretty good job.

[Nathan Zylbersztejn @nathan_zyl]

Good software is very much alive. You can vibe clone anything in a weekend, but it’ll just suck. Example: there are thousands of dictation apps 99.9% suck and are unusable.

[@levelsio @levelsio]

Software is dead

It's now the era of personal software and everyone just builds what they need whenever they need it I think

[Nathan Zylbersztejn @nathan_zyl]

@levelsio I have been a host for 12 years, 100s of stays, got only 2 reviews removed and that was not easy. Properties with low ratings get no booking and leave the platform, so only 4+ remain.

[Nathan Zylbersztejn @nathan_zyl]

@karpathy You’re at Anthropic. Can you get Claude to write is STE by default so we don’t need to?

[Nathan Zylbersztejn @nathan_zyl]

@mattshumer_ for those who don't get how to use them

[Nathan Zylbersztejn @nathan_zyl]

@elshayib_ People just don't know how to use them. Prompting a Qwen 3 0.6B and expecting a useful answer is stupid; however, you can leverage its language knowledge and probabilities for specific tasks, such as correcting ASR transcription.

[Nathan Zylbersztejn @nathan_zyl]

It is a vivid topic in France after a novel shortlisted for the Goncourt Prize (the Oscars of French literature) was removed over suspicions that it was written with AI.
True or not, if it made its way to the Goncourt and to the NYT before , it must have been genuine to a certain extent.
https://t.co/DKaqWAHnwe

[Nathan Zylbersztejn @nathan_zyl]

@SamuelZengML Any plan to integrate your models with Apple Metal and FluidAudio?

[Nathan Zylbersztejn @nathan_zyl]

@adin_ron @typesafeai Nice demo!

[Nathan Zylbersztejn @nathan_zyl]

@_catwu, congrats on the latest editable Artifacts. Previously, they were hard to refine, make slope-free, and impossible to share as finished documents.

But at this point, you're rebuilding Notion within Artifacts. Integrating more closely with Notion, where we store and share our knowledge, could be more productive. E.g we could draft TDs directly in Notion (we have a TDs db) with Claude's help instead of starting in Artifacts, copying into Notion, refining manually, and doing back-and-forth with Claude and Notion MCP.

[Nathan Zylbersztejn @nathan_zyl]

@RhysSullivan Because they're trying to solve ALL problems with a single magical prompt, ignoring decades of classic NLP work that could actually help.

[Nathan Zylbersztejn @nathan_zyl]

I started looking at dictation tools a couple of years ago because my hand pain started growing due to genetic peripheral neuropathies. It became harder to get to the end of the day and the end of the week. I thought I had only a few years left to work.

I started with Apple Dictation, and realized it wasn't as terrible as I expected. Lots of corrections to do, but much less typing at the end of the day. The hardest part was shifting from writing to speaking, and even today I am more comfortable writing things that require careful thinking, but I can ramble all day long with AI agents, and they get me just fine.

I tried Wisprflow, of course, and it's a decent product, but sending my all-day-long ramblings to a remote server isn't for me.

I think whatever tool you're using, it's hard not to put the final touches on a dictation before sending it to someone. I'm fine typing when it matters.

What was really frustrating for me was all the repetitive changes, such as adding a PR link when asking for a review, adding a mention in a Slack message, or correcting dates and amounts. Most products try to be one-size-fits-all, but we're all marginally different and ideally need to customize the "last mile" to our personal workflows.

I hope your hands get better...

[Nathan Zylbersztejn @nathan_zyl]

I built ParrotFlow: local dictation for macOS.

Shape your dictation with rules, code, and prompts. Small local models choose the right terms from your vocabulary in context. Built-in cleanup keeps your wording, without an LLM rewrite.

35-second demo ↓

[Nathan Zylbersztejn @nathan_zyl]

@bcherny Candid question: were all 16 PRs reviewed by someone? How do you handle the PR volume?

[Nathan Zylbersztejn @nathan_zyl]

@HedgieMarkets When AMZN offers to finance your purchase with a Klarna equivalent, is it circular too?

[Nathan Zylbersztejn @nathan_zyl]

@HedgieMarkets How do you explain huge compute deals such as MSFT renting Mistral’s compute if clients are not showing up?

[Nathan Zylbersztejn @nathan_zyl]

@RaoulGMI Is it before or after Bitcoin is worth 1 million? 😂