imranq

Everyone saying you could replace Jev or decision type models with an LLM with bolted schema output constraining are missing the point completely. Its about extreme speed and cost effectiveness with high quality, neither of which you are going to get with LLMs even with these KV-cache tricks

The main reason to learn something is actually being able to communicate in the language of that subject. There are complex ideas in math that cannot be easily captured by the language of other fields. Pepole who don't study math cannot even understand what a worthwhile goal in math even is or how it could be useful to other fields.

You can't simply prompt a model to be "better" when "better" isnt even properly defined

This guy just seems a bit salty

> Officials said some users expected Maven to flag stale records or contradictions in the intelligence assembled for potential targets, though it is not clear why they thought the system would do that

This is what happens when people who don't understand AI and its limitations embrace it wholeheartedly, thinking it's going to behave like The ultimate analyst. Except it's not. Its got blind spots in more places than anyone knows. And not to mention the system itself has multiple areas of failure.

Based on my admittedly limited research, it seems like you should use Laya for much more deterministic tasks where you have some training data. It won't be as good as Jev for zero shot cases.

Note that this proof while impressive does not add any value to mathematics as a human pursuit. But it does show we can throw these LLM beasts at much gnarlier problems than we could have imagined previously. Maybe even formally verify papers the day they are posted?

I'd love to see an e2e compiler or OS kernel verification or Full-stack chip design with formal equivalence checking at each stage that would be pretty cool.

What else is interesting is how they staged this problem : (a) maintain an explicit DAG/roadmap of sub-goals rather than one flat prompt, (b) separate statements from proofs so many agents can work on different nodes without stepping on each other, (c) keep a natural-language index alongside the formal one so search/reuse works... I feel like this is the future of long horizon agents and how you can do work that's making the most of every agent. This approach will likely be baked into the next versions of coding harnesses

Transcribed letter (from chatgpt):

Dear Mr.,

In your letter of December 14, you ask me whether mathematics is a worthwhile endeavor. Mathematics absolutely is not a worthwhile endeavor! I can imagine only three possible reasons for studying mathematics:

(1) Entertainment. (2) To promote technological progress. (3) To promote one’s own career.

Reason (1) is harmless; but one doesn’t refer to entertainment as a “worthwhile endeavor”; it’s just a pastime.

As for reason (2), modern technology is the worst thing that ever happened to the world, and to promote its progress is nothing short of criminal. Read Our Final Century, by Martin Rees; or read “Why the Future Doesn’t Need Us,” an article by Bill Joy that appeared in Wired magazine, April, 2000.

As for reason (3), to choose mathematics as one’s career would be very risky. Not long ago I received a letter from a gravestone sculptor whose craft had been rendered obsolete by a laser-guided machine that carved gravestones automatically. He was in his forties, unable to find work, and depressed. Maybe you think that can’t happen to you if you become a mathematician. But you’re wrong. Martin Rees, Bill Joy, and many other distinguished experts believe that within a few decades machines will be developed that will be more intelligent than even the smartest human beings. When that happens, people will be obsolete; they will be phased out and will be replaced by computers and robots. This will happen to everyone eventually, but mathematicians probably will become obsolete before anyone else does, because mathematics is more amenable to computerization than any other field. Hence, computers probably will surpass humans in mathematics before they do so in other areas.

Sincerely yours, Ted Kaczynski

Personally, I found that people are impressed with AI for subjects they simply don't understand, and when an expert works with AI, they're often underwhelmed. So while we're seeing some mathematics being automated and proved, I think there's a long way to go before we automate proofs with meaningful human value.

Google doesnt even use its own Go build system internally. Its all blaze / bazel, so they are not even taking advantage of the so called compiler feedback of Go. Also if languages are to be designed for agents not humans, its not clear whether the verbosity of Go will help agents at all

Okay so the OP is saying that since Macbook Neo has the same hardware as iphone, but not locked down, so why is iphone locked down. They say its because of the app store profits.

Sure App store is not to be understated, but I'd add our phones include way way more personal information than a laptop like NFC for credit cards, personal photos, and all biometric and contact information. Not to mention cellular network connection and generally forms as a soft form of identity. None of these apply to a laptop. So form factor does matter.

BUT even if we unlocked the iPhone, the desire for 'MacOS on iPhone' is actually the wrong thing to ask for. Pete Steinberger had in this interview (https://www.youtube.com/watch?v=AcwK1Uuwc0U&t=1182) that UI is basically the wrong paradigm in a world where agents should do tasks for us in milliseconds. We should be able run any local services from our phone like grabbing

Good news is we already have this via terminal apps in Android. Now what's left is the ability for agents to run on your device and basically accomplish tasks for you

Amazing write up and i wish more people showed the process for discovery which is often even more interesting than the result itself

Still the result is really interesting being able to stack abstract reasoning and get better performance and the heat maps to show the prob results

The academic literature seems to be catching up:

- *[SOLAR / DUS (Kim et al., 2023)](https://arxiv.org/abs/2312.15166)* — duplicated transformer layers to build a 10.7B model that outperformed 30B parameter baselines.

- *[The Curse of Depth (2025)](https://arxiv.org/abs/2502.05795)* — explains why this works: Pre-LN causes deep transformer layers to converge toward identity functions, meaning middle layers are where real computation happens, and duplicating them concentrates that capacity.

- *[Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach (Geiping et al., NeurIPS 2025)](https://arxiv.org/abs/2502.05171)* — takes the idea to its logical conclusion: a model trained with a single recurrent block repeated at inference time, scaling reasoning depth without adding parameters.

I think it might be 2020 when the M1 was released since I remember i had bought a mac book in 2019 and it was still intel

I really like Clawdbots safety gloves off approach - no handholding or just saying yes to every permission. I just message my server using telegram and it does research for me, organizes my notes, and builds small apps on the fly to help with learning

Some great life lessons here, but also some I don't agree with:

- The lazy person works twice as hard. Often I found you can save a lot of time just trying to the minimal possible and gain a lot of insights of why something is minimal vs not

-The opinion of the person who rarely offers it is listened to more closely. I found the opposite to be true, those who don't offer their thoughts frequently are often dismissed when they do want to share something

Anyway, many of the points are great.. I would also add to keep a journal and write down what was meaningful throughout the day.. you will find time passing by with more quality since you know what the take and what to avoid

Just because it is in C, doesn't mean you will get C like performance. Just look at the benchmarks, it is 8x slower than just using PyTorch... while I get its cool to use LLMs to generate code at this level, getting super high performing optimized code is very much out of the domain of current frontier LLMs

I really liked the approach of getting new topics to research via embeddings, trails, and claude code, but often what will this give you outside of novelty?

Hey HN! I built an explorer for all NeurIPS 2025 Main Conference and workshop papers with reviews, scores, and code links.

But the unique feature: AI-generated "explainers" that break down complex papers with interactive visualizations. Example: https://neurips2025.pages.dev/explainers/linear_attention/

It explains why attention is hard to optimize, shows the math with interactive demos, and includes critical analysis of limitations.

The explainers are generated using Gemini3 to parse papers and create:

- Interactive visualizations

- Step-by-step mathematical walkthroughs

- Critical analysis sections

- "What would convince me?" sections

Tech stack: OpenReview API, Gemini API for explainer generation, static hosting on Cloudflare Pages for speed.

I'm planning to generate explainers for more papers based on what people find interesting, so any feedback would be amazing

The claims in this paper don't make sense. There is no proof that anything has been decompressed

There is one: https://pricepertoken.com/

Turning it off and then on again works in a lot of surprising places

People basically want a life coach, someone by their side who can tell them what the best next thing to do is at any given moment. Everything else are just approximation of that ideal.

The author's .txt file works because its simplicity forces a daily ritual of self-coaching. The tool demands that the user manually review, prioritize, and decide what matters. There are no features to hide behind, only the discipline of the process itself.

The impulse to use complex apps or build custom scripts is the attempt to engineer a better coach. We try to automate the prioritization and reminders, hoping the system can do the coaching for us.

The great trap, of course, is when we fall in love with engineering the system instead of doing the work. This turns productivity into a sophisticated form of procrastination.

Ultimately, the best system is the one that removes the most friction between decision and action. For the author, that meant stripping away everything but the list itself.

While I agree with the author's vision for a more human-centric AI, I think we're closer to that than the article suggests. The core issue is that the default behavior is what's being criticized. The instruction-following capabilities of modern models mean we can already build these Socratic, guiding systems by creating specific system prompts and tools (like MCP servers). The real challenge isn't technical feasibility, but rather shifting the product design philosophy away from 'magic button' solutions toward these more collaborative, and ultimately more effective, workflows

My two takeaways is you build 1) Having a precise vision of what you want to achieve 2) Being able to control / steer AI towards that vision

Teams that can do both of these things, especially #1 will move much faster. Even if they are wrong its better than vague ideas that get applause but not customers

This was a bit hard to read. It would be good to have a narrative structure and more clear explanation of concepts.

I'd love to host my own LLMs but I keep getting held back from the quality and affordability of Cloud LLMs. Why go local unless there's private data involved?

I like the idea of removing quadratic scaling for attention, this paper has thin experimental support. No real tasks tested beyond perplexity. Nothing on reasoning, retrieval QA, or summarization quality. Even in perplexity the gains are marginal.

However it removes attention so I think its worth watching that space of non-attention models

The powers of kerning are great indeed

Nice idea!

I think it would be better if it was just context and not connected to any model. Think of one place where you can hook in your drive folder, GitHub, etc. and have it produce the best context for the task you want to achieve. Then users can copy that to their model or workflow of choice

Could you pop your code into an LLM and ask it to write comments for you? I'm not sure how accurate it is though