thorum

Isn’t the goal of Study Mode exactly that, though? Instead of handing you the answers, it tries to guide you through answering it on your own; to teach the process.

Most people don’t know how to do this.

Unintended side effect, UK MPs can now watch as much porn as they want with plausible deniability.

The long term is unlimited access to local LLMs that are better than 2025’s best cloud models and good enough for 99% of your needs, and limited access to cloud models for when you need to bring more intelligence to bear on a problem.

LLMs will become more efficient, GPUs, memory and storage will continue to become cheaper and more commonplace. We’re just in the awkward early days where things are still being figured out.

This is independent research, not by the Google team that won gold, showing the public model can achieve gold results through clever prompting.

It implies that training on synthetic data will always shift the model’s behavior in unpredictable ways. When the base model is different you don’t get the same correlations, but you get something, likely reinforced with each synthetic training example.

The greater variance of real world data might avoid this effect.

I recommend litellm if you’re writing Python code, since it handles provider differences for you through a common interface:

https://docs.litellm.ai/

Beyond a basic understanding of how LLMs work, I find most LLM news fits into one of these categories:

- Someone made a slightly different tool for using LLMs (may or may not be useful depending on whether existing tools meet your needs)

- Someone made a model that is incrementally better at something, beating the previous state-of-the-art by a few % points on one benchmark or another (interesting to keep an eye on, but remember that this happens all the time and this new model will be outdated in a few months - probably no one will care about Kimi-K2 or GPT 4.1 by next January)

I think most people can comfortably ignore that kind of news and it wouldn’t matter.

On the other hand, some LLM news is:

- Someone figured out how to give a model entirely new capabilities.

Examples: RL and chain of thought. Coding agents that actually sort of work now. Computer Use. True end-to-end multimodal modals. Intelligent tool use.

Most people probably should be paying attention to those developments (and trying to look forward to what’s coming next). But the big capability leaps are rare and exciting enough that a cursory skim of HN posts with >500 points should keep you up-to-date.

I’d argue that, as with other tech skills, the best way to develop your understanding of LLMs and their capabilities is not through blogs or videos etc. It’s to build something. Experience for yourself what the tools are capable of, what does and doesn’t work, what is directly useful to your own work, etc.

I understand the idea. My position is that this is a largely speculative claim from people who have not spent much time seriously applying agents for spreadsheet or video editing work (since those agents didn’t even exist until now).

“Getting something almost right, no matter how close, can often be worse than not doing it at all” - true with human employees and with low quality agents, but not necessarily true with expert humans using high quality agents. The cost to throw a job at an agent and see what happens is so small that in actual practice, the experience is very different and most people don’t realize this yet.

People say this, but in my experience it’s not true.

1) The cognitive burden is much lower when the AI can correctly do 90% of the work. Yes, the remaining 10% still takes effort, but your mind has more space for it.

2) For experts who have a clear mental model of the task requirements, it’s generally less effort to fix an almost-correct solution than to invent the entire thing from scratch. The “starting cost” in mental energy to go from a blank page/empty spreadsheet to something useful is significant. (I limit this to experts because I do think you have to have a strong mental framework you can immediately slot the AI output into, in order to be able to quickly spot errors.)

3) Even when the LLM gets it totally wrong, I’ve actually had experiences where a clearly flawed output was still a useful starting point, especially when I’m tired or busy. It nerd-snipes my brain from “I need another cup of coffee before I can even begin thinking about this” to “no you idiot, that’s not how it should be done at all, do this instead…”

> The leaked responses show clear signs of being real conversations: they start with contextually appropriate replies, sometimes reference the original user question, appear in various languages, and maintain coherent conversational flow. This pattern is inconsistent with random model hallucinations but matches exactly what you'd expect from misdirected user sessions.

A model like GPT-4o can hallucinated responses that are indistinguishable from real user interactions. This is easy to confirm for yourself: just ask it to make one up.

I’m certainly willing to believe OpenAI leaks real user messages, but this is not proof of that claim.

The interesting part of context engineering (the actual engineering part) is figuring out how to gather the information the LLM needs to do a task correctly from your system. For example, the secret sauce of GitHub Copilot is how it decides what parts of your codebase to show the LLM. This is surprisingly hard when you need something other than simple RAG. In many cases the data source you need doesn’t exist and you have to build it.

The prompt engineering side of the problem (how you structure your prompt) is trivial by comparison and will become less and less relevant as frontier models improve.

It’s very, very hard to remove things from the training data and be sure there is zero leakage.

Another idea would be to use, for example, a 2024 state of the art model to try to predict discoveries or events from 2025.

Interesting project! I’m a little surprised that Claude is willing to call these functions. The demo screenshot is downloading a public domain work, I wonder if it would also happily go along with requests for Harry Potter or other copyrighted material?

Some ideas are too complex to explain accurately in simple terms.

You can give someone a simple explanation of quantum chromodynamics and have them walk away feeling like they learned something, but only by glossing over or misrepresenting critical details. You’d basically just be lying to them.

I’ve built several apps on yjs and highly recommend it. My only complaint is that storing user data as a CRDT isn’t great for being able to inspect or query the user data server-side (or outside the application). You have to load all the user’s data into memory via the yjs library before you can work with any part of it. There are major benefits to CRDTs but I don’t think this trade-off is worth it for all projects.

This GitHub issue says 6-7 GB VRAM: https://github.com/resemble-ai/chatterbox/issues/44

But if the model is any good someone will probably find a way to optimize it to run on even less.

Edit: Got it running on an old Nvidia 2060, I'm seeing ~5 GB VRAM peak.

We’ll know AGI has arrived when it can figure out Python dependency conflicts

Try AI Studio if you haven’t already: https://aistudio.google.com/

Constraints make the specific goal (moon landing) harder, but force technological development. If landing on the moon had been 'easy' with existing tech and not required that massive investment of resources, progress on everything else is delayed. Material science, engines, batteries, solar, radios, integrated circuits, even PCs and smartphones - sure it would all have happened eventually but the key innovations were made when they were because of constraints.

With Jules, I almost always end up making significant changes before approving the PR. So “successful merge” is not great indicator of how well the model did in my case. I’ve merged PRs that were initially terrible after going in and fixing all the mistakes.

> To actually get to the bottom of things: I think most normal folks are concerned more about getting by and making decent money in “the age of AI” than they are about being brilliant whizkid prodigies coming up with original ideas. A lot of those end up being poor anyway. But the desire to live a quality life is a more universal thing. No amount of “mind training” will help here. Just steer clear of paths that AI can dominate (they’re expanding), and failing that, use it to your advantage as best as you can.

In my experience most people working on this are attempting variations on a handful of generic ideas. Most AI startups are fairly uninspired “XYZ but with AI chat” type things or ideas that have no staying power because Claude 7 will one-shot the whole product with a prompt. Succeeding here in the long run means doing something truly different and new and interesting and that’s what the linked articles are about.

Thanks for the recommendations and I agree completely.

There's some hope, we can get better at this even in small ways by carving out pockets of time away from phones and notifications and other external distractions and inputs.

If you're interested in preserving your ability to think for yourself in the age of AI, I recommend Henrik Karlsson's blog Escaping Flatland. While not directly about AI, his articles "Cultivating a state of mind where new ideas are born" [1] and "Childhoods of exceptional people" [2] explore similar themes of how to train your mind to have original ideas and learn to solve problems on your own.

[1] https://www.henrikkarlsson.xyz/p/good-ideas

[2] https://www.henrikkarlsson.xyz/p/childhoods

https://www.henrikkarlsson.xyz/archive?sort=top

Makes sense, I could see the human touch on the article too, so I figured it was something like that.

Humorous that this article has a strong AI writing smell - the author should publish the prompts they used!

Good article. Agree that general unreliability will continue to be an issue since it's fundamental to how LLMs work. However, it would surprise me if there was still a significant gap between single-turn and multi-turn performance in 18 months. Judging by improvements in the last few frontier model releases, I think the top AI labs have finally figured out how to train for multi-turn and agentic capabilities (likely RL) and just need to scale this up.

Good points. I suspect that o3 is able to reason more deeply about different paths through a codebase than earlier models, though, which might make it better at this kind of work in particular.

Sorry, I added links! Just a week ago someone built a system that used o3 to find novel zero days in the Linux kernel’s SMB implementation.

My takeaway - from this article, from Google’s AlphaEvolve [1], and the recent announcement about o3 finding a zero day in the Linux kernel [2] - is that Gemini Pro 2.5 and o3 in particular have reached a new level of capability where these ideas that were tried unsuccessfully with other models, suddenly just work.

[1] https://deepmind.google/discover/blog/alphaevolve-a-gemini-p...

[2] https://sean.heelan.io/2025/05/22/how-i-used-o3-to-find-cve-...

Apple makes webapps/PWAs hard because they want you to make a native app instead.