exizt88

I don't understand how these benchmarks get things so wrong. Nova-3 is by far the best model on the market in WER. Azure's STT is atrocious. And yet every benchmark rates Azure's STT higher than Nova-3's.

I'd give a lot to see the WA group chat between the Instinct founder and their VCs.

Water consumption of AI is a non-issue. Why even spend research and policymaking resources on this?

As someone who's been doing AI call centres for a while, I'm not surprised. It's really, really hard to build voice AI that works. There are no open source solutions that work out of the box.

But the biggest problem is ASR. WER is still atrocious even with SOTA models. When you add drug names and regional accents, it's a recipe for disaster.

> For AI2027 to be real, the money has to come from somewhere to carry on building the economy. If >10% of the workers suddenly become unemployed, and the rest taking paycuts, then money supply dries up. (unless central banks do something, but then that can be highly inflationary)

Why would it be inflationary?

> We are seeing improvements with each model release these days but it’s clear that the improvements are getting smaller and smaller.

This is obviously untrue, both with GPT-5.4, and Claude Fable as examples in the last 6 months.

You guys absolutely rock. You've built an industry standard with Promptfoo. Keeping it open source was the right choice.

The "Less Growth for Europe" party strikes again.

The reason for divergence is actually much simpler. NASDAQ 100 includes data center builders, Morgan Stanley software index doesn't. Stock market is going down across the board if you exclude data center construction.

Marc Maron has coined the perfect description for sanctimonious posts like these: "annoying Americans into fascism".

Recombine | Principal Software Engineer | Remote (UK/EU) | Full-time | https://recombine.ai |

Recombine builds advanced conversational AI agents for Europe's banks, fintechs and marketplace companies.

We're hiring a Senior Solution Engineer, a Principal Software Engineer and a Senior Infrastructure Engineer.

Tech we use: TypeScript, Node.js, Python, Terraform. We are cloud-agnostic.

We don’t sponsor visas at the moment.

See all of our opportunities here: https://recombine.ai/careers/

Is their CEO comp not in line with the market?

If the accusers had evidence, they would have surely provided it to the defendants in the defamation case.

> OPERATORS: No precedence, executed left to right, parenthesize as desired. 2+3*10 yields 50.

How do you even come up with this?

One of my favorite things about Curtis Yarvin is that one of his main ideas is laughably easy to prove wrong. He says that whenever we go into the past, the previous generation seems to be more "right-wing" than the current one. Which is almost comically wrong. For example, you get Victorian Era being much more conservative than its predecessor, Georgian Era. Same goes for Christian Roman Empire vs Pagan Roman Empire, Nazi Germany vs Weimar Germany, etc etc. There are literally dozens of examples. It seems that Yarvin genuinely doesn't know any of these high school-level facts.

> chatgpt-4o-latest, but it's bad because they continuously update it

Version explicitly marked as "latest" being continuously updated it? Crazy.

For conversational AI, the most significant part is GPT-4.1 mini being 2x faster than GPT-4o at basically the same reasoning capabilities.

Unfortunately, based on my experience, GPT-3.5 is simply not up to snuff in phone conversations in terms of style and reasoning. And GPT-4T latency is far too large for real-time conversation with straightforward STT-TextGen-TTS pipeline.

Carbon Monoxide is odorless, but it’s effects are still real and well-understood.

If your goal is to produce a lot of value and you don’t care about others capturing it, then it may actually be a good way to go, especially with the non-profit setup.

I bet "new advanced AI research team" at Microsoft is going to be underwhelming for many, but really, it should be eye-opening. This is what startups, especially VC-backed capital-intensive AI startups, usually are.

Anyone who works with text generation will tell you that GPT-4 is far, far beyond anything anyone else has put out for general purpose text gen. The benchmarks don’t really tell you the whole picture. It’s impossible to prompt other models for anything as complex as what GPT-4 can do, both semantically and stylistically.

How do you know that? Maybe they wanted to ship AI products at an unprecedented speed at the most prestigious AI company in the world.

The weights aren’t MIT-licensed, so this is not usable in commercial applications, right?

That’s exactly the point of Replit. It removes a lot of the barriers that are keeping people from using programming in their daily lives.

The goal of personal computing was to make computers, with all their power, available to everyone. In that Replit is much more of a continuation of that work than its antagonist.

The money quote in the second video is at 11:35: “It’s very spicy”

Yeah, but your customer acquisition cost becomes your overhead. And it can easily take up 30-50% of your unit economics. Anyone who's tried building SaaS for creators knows that it's a very tough crowd to sell to.

Anyone who has tried to build a pay-by-work model for digital creators knows that it’s simply an awful business model. There just isn’t that much demand for creator work. And there are no “whales” among these customers to compensate. No one’s buying 10 copies of a math explainer video.

That’s where directives for enterprise contracts usually come from. I’m sure they won’t even talk to anyone not willing to pay $100k+ per year. Salesforce’s AI Cloud starts at $365k a year.