thorum

It works on iOS 18. Safari’s audio support was pretty rough on earlier versions, especially <17, but I didn’t need to support those for my project.

The solution I found after approximately two months of struggling with this problem: you have to generate an audio file that is a few seconds of silence, play it on a loop, and play it at the same time as the actual audio file you want to play (via separate audio elements, or an AudioContext). Specifically I believe you need to make sure the silence is “playing” at track boundaries for the real audio, so there is never a single moment where your webapp stops playing audio.

Google’s ability to offer inference for free is a massive competitive advantage vs everyone else:

> Is Jules free of charge?

> Yes, for now, Jules is free of charge. Jules is in beta and available without payment while we learn from usage. In the future, we expect to introduce pricing, but our focus right now is improving the developer experience.

https://jules-documentation.web.app/faq

GPT 4.1 and 4o score very low on the Aider coding benchmark. You only start to get acceptable results with models that score 70%+ in my experience. Even then, don't expect it to do anything complex without a lot of hand-holding. You start to get a sense for what works and what doesn't.

https://aider.chat/docs/leaderboards/

I think realistic simulations of social interactions would help many people - by providing a low pressure environment for socially awkward people to practice the skills you mentioned, before putting themselves out there in groups etc. - but realism is an absolute requirement. No current AI will do that for you, they are not good at roleplaying realistic humans, and they’re too nice.

If you want a “practice date” to give you useful training that transfers to real life, your conversation partner needs to react to you just like a human would. Including negative reactions (not liking you, thinking you’re weird, faking a call from a friend to escape mid-conversation, and so on). But if you could make an AI that did all that, and it could also give you the sort of actionable feedback that is hard to get from humans, I think that has potential.

The honest version of this feature is that Gemini will act as your personal assistant and communicate on your behalf, by sending emails from Gemini with the required information. It never at any point pretends to be you.

Instead of: “Hey garry, my daughter woke up with the flu so I won't make it in today -Pete”

It would be: “Garry, Pete’s daughter woke up with the flu so he won’t make it in today. -Gemini”

If you think the person you’re trying to communicate with would be offended by this (very likely in many cases!), then you probably shouldn’t be using AI to communicate with them in the first place.

> Very good knowledge of history, science, technology, everything

All of English Wikipedia is about 100 GB. A local LLM that can run searches against a local Kiwix server would work. Local LLMs can’t really do this correctly yet, but it’s not inconceivable in 1-2 years.

I think that’s an opportunity, not a problem. If prompt + hint generates a verifiable solution then you can build systems that propose hints, either randomly or by exploring a search space, and keep trying combinations until you hit on something that works.

There’s no way to reward truthfulness that doesn’t also reward learning to lie better and not get caught.

They're listing the use cases where a 10-20% hallucination rate isn't a big deal. They can't advertise the truly useful applications of the tech because it's not reliable enough (yet).

But they can make it more reliable over time by deploying the product now and training on user data:

> Google collects your chats (including recordings of your Gemini Live interactions), what you share with Gemini Apps (like files, images, and screens), related product usage information, your feedback, and info about your location. (...)

> Google uses this data, consistent with our Privacy Policy, to provide, improve, and develop Google products and services and machine-learning technologies, including Google’s enterprise products such as Google Cloud.

https://support.google.com/gemini/answer/13594961?hl=en

The OpenAI library is used pretty much everywhere in the LLM world, every serious AI provider has an OpenAI-compatible API, or you can use something like OpenRouter which is also OpenAI-compatible. Standardization is good.

In the age of local LLMs I’d like to see a personal recommendation system that doesn’t care about being scalable and efficient. Why can’t I write a prompt that describes exactly what I’m looking for in detail and then let my GPU run for a week until it finds something that matches?

It’s kind of like ChatGPT writing, where it can easily fool people who see it for the first time, but after a while you start to recognize the common patterns.

Why does this AI-generated article have 400 upvotes? The conversation here is valuable, but surely we could find a human article to have it under instead?

More than just editing, I’d say. Large sections are clearly 100% AI.

Hm, you’re right. I’m thinking a nearby Raspberry Pi with a camera facing the box. It uses an LLM to identify the box contents and send update messages to Kasia each morning. I’m surprised they didn’t think of this already.

There’s also an FAQ discussion about the language choice: https://github.com/microsoft/typescript-go/discussions/411

Isn’t that moving the goalposts? The claim was made that it’s impossible to detect AI training runs and investigate what’s going on or take regulatory action. In fact, it is very possible.

AIME has significant problems: https://x.com/DimitrisPapail/status/1888325914603516214

> Problems near identical to the test set can be found online.

The key thing here is a simple, reliable formula to train a 1B model on a specific task and get strong performance. That didn’t really exist before. Edge devices are about to get a lot smarter.

Potentially a good source of startup ideas: Look where AI is being underutilized vs its economic potential and go there.

This seems like a problem that will quickly fall to the new reinforcement learning methods introduced by DeepSeek. Just build a system to synthetically render a few million pages of insanely complex, hard-to-parse documents with different layouts along with a JSON description of what the correct OCR should be, mix in some human annotated datasets, then do RL against a verifier that insists on 100% accuracy.

Not really. They’re successful because they created one of the most interesting products in human history, not because they have any idea how to brand it.

I was getting pretty hyped about the potential for GRPO in my own projects until you said 20 minutes for a single training step with batch size 1! Is that likely to improve?

> “It is (relatively) easy to copy something that you know works,” Altman tweeted. “It is extremely hard to do something new, risky, and difficult when you don’t know if it will work.”

The humor/hypocrisy of the situation aside, it does seem to be true that OpenAI is consistently the one coming up with new ideas first (GPT 4, o1, 4o-style multimodality, voice chat, DALL-E, …) and then other companies reproduce their work, and get more credit because they actually publish the research.

Unfortunately for them it’s challenging to profit in the long term from being first in this space and the time it takes for each new idea to be reproduced is getting shorter.

Yes, “take this clever code written by a smart human and convert it for WASM” is certainly less impressive than “write clever code from scratch” (and reassuring if you’re worried about losing your job to this thing).

That said, translating good code to another language or environment is extremely useful. There’s a lot of low hanging fruit where there’s, for example, an existing high quality library is written for Python or C# or something, and an LLM can automatically convert it to optimized Rust / TypeScript / your language of choice.

It’s great, but it only works for problems where there is exactly one correct solution and it’s possible to automatically verify the solution - like math and programming. So far these reasoning models have not shown much transfer learning of reasoning to other domains, and are often worse at non-math/code tasks than standard models.

> The point is to make it difficult.

Does it, though?

The practical problem I see is that unless US AI labs have perfect security (against both cyber attacks and physical espionage), which they don’t, there is no way to prevent foreign intelligence agencies from just stealing the weights whenever they want.

Connecting people to other people, to life changing art, places and things that they end up loving and wouldn’t know about otherwise? That has to be one of the best uses for technology. I’d like to see more of it.

I think you and others here are focusing on the stereotypical “influencer” faking authenticity for views but there are literally millions of human beings posting on TikTok about all kinds of things. A lot of them are pretty cool. Just click “not interested” on influencers and click like on the stuff you want to see instead.