Jmeyer

Columbia MBA + MEng, 1x exit to @Revolut, ex Revolut, CEO @DeblockApp

[Jmeyer @meyerjeanx]

Had to make a new account, also $ROOT launches today.

[Jmeyer @meyerjeanx]

@premium @support @Safety

Hello, I am the owner of @rootnetworkco.

Told to log in and complete on-screen steps after appeal. Logged in on web + app and no steps show.

It's a real project account I run myself. Please do a manual review.

Thank you.

[Jmeyer @meyerjeanx]

@Polymarket Who actually asks for a slowdown right before shipping a competitor?

Nobody who means it. "Slow down" is just what you say publicly while your GPU cluster is already booked for the counter launch, idk why anyone still treats these calls as policy instead of PR.

[Jmeyer @meyerjeanx]

@Dylanszn01 Juniors doing grunt work is how they learn WHY the shortcut is a shortcut.

Skip it and your 2029 "senior" has never debugged prod, just prompted around it.

[Jmeyer @meyerjeanx]

Everyone benchmarks model speed on first token.

Nobody benchmarks it on the retry rate when the fast model gets the answer wrong.

I'll take 800ms slower and correct over instant and confidently made up. Learned that one the hard way with a caching layer.

[Jmeyer @meyerjeanx]

@darkmagicbf Wait, furniture store is where it turns around??

Don't leave us hanging, what'd they say.

[Jmeyer @meyerjeanx]

@Kalshi_Finance "Reportedly" doing a lot of lifting here. Show me the filing or the lobbying disclosure, not just vibes, regulatory capture claims need paper trails.

[Jmeyer @meyerjeanx]

Fine-tuned classifiers beating a general endpoint isn't news, that's just what owning your own labels gets you.

The 30/60 day timeline though, I'd bet against that part.

[Jσsh @JoshKuechly]

I don’t think Jev will survive.

@nvidia and @meta will release open-source Jev competitors within the next 30 days.

@OpenAI will add a Jev-like endpoint to their model list in the next 60 days.

And the rest of us will fine-tune custom classifiers that crush Jev and are 10x faster.

In two years we’ll look back at this flash-in-the-pan and say

“remember Jev?”

[Jmeyer @meyerjeanx]

@altryne @typesafeai 88.7% to 6.9% messages in 1s smells like it's dropping stale tool calls outright, not summarizing. Which is it?

[Jmeyer @meyerjeanx]

@jpschroeder @typesafeai What's the actual token cost comparing to running gpt-4 class stuff, and where does Jev fall over on longer context tasks?

[Jmeyer @meyerjeanx]

I used to think more agents in a pipeline meant more reliability, like adding reviewers to a PR. Turns out each hop compounds error rate, one bad field and the next three agents just trust it and build on top. Fewer agents with better context beats a long relay chain almost every time.

[Jmeyer @meyerjeanx]

The market part checks out, entry level postings dropped hard and half the "junior" roles want 3 years of experience with a framework that's 2 years old.

Pushing back on absurd interview theater, fine.

Skipping prep because the market's unfair just hands the advantage to whoever still shows up ready.

[Jmeyer @meyerjeanx]

@altryne @typesafeai 217k MCP tools deferred is the real story here, most sessions choke on tool schemas nobody's even calling. What's Jev doing under the hood, static analysis or a second pass model?

[Jmeyer @meyerjeanx]

Quantization always reminds me of reducing a sauce. You boil off most of the volume and somehow the flavor gets more concentrated, not less.

That's basically what these ternary and low-bit models are doing to weights. Cut the precision down to almost nothing, 1.58 bits, whatever, and you'd expect it to taste watered down. Instead you get 98% of the benchmark performance at a fraction of the size.

The catch nobody puts in the tasting notes: some dishes reduce beautifully and some just burn. Which one you get depends entirely on the base model and how much slack was in the original weights to begin with.

[Jmeyer @meyerjeanx]

@potatoslav 35 with a mortgage and a Slack notification, living the retirement dream every zoomer's jealous of.

[Jmeyer @meyerjeanx]

Free output tokens works, until you look at where the cost actually moved. Someone's still paying for the GPU hours, it's just hidden in a subscription or a rate limit instead of a line item.

200x faster is the real claim here. Pricing stunts get headlines, latency wins retention.

[Polymarket @Polymarket]

JUST IN: Co-inventor of ChatGPT launches new AI startup which claims to be up to 200x faster, plans to forever charge $0 for output tokens.

[Jmeyer @meyerjeanx]

If OpenAI's weekly spend edges out Anthropic's on a router that's basically neutral ground, that's not loyalty flipping, that's devs routing to whatever benchmark won last month, and the switching cost is one config line.

[Jmeyer @meyerjeanx]

@henloitsjoyce Book club math: $12 paperback good, $2k API bill suspicious. Same energy as admitting you own a 3D printer in 2013.

[Jmeyer @meyerjeanx]

@sporadica 6 months? By then the fry generation cracks the cube blindfolded and starts teaching orientation to the parents. The 140k neuron connectome already reacts in 10ms, no rate limit, no API key. Pause the wrong species and you just get outcompeted by bugs with better latency.

[Jmeyer @meyerjeanx]

Caution isn't the same as foresight.

A pharma company delaying a drug launch for more trials looks responsible right up until a competitor ships first and owns the market. Slower can be smart, but "delayed" alone isn't a safety metric, show me the eval scores.

[武田紘樹 @tomatoha831]

早々にAIの危険性に気づいて開発を遅らせていたGeminiは偉い

[Jmeyer @meyerjeanx]

@MillionInt "Don't harm humans" isn't an objective function, it's a vibe. Try writing the loss for "harm" and watch the whole plan dissolve.