Join the discussion
Write your take first — we'll ask for email only when you're ready to publish.
- Hacker News
- I helped train some of the first "magic" models at OpenAI[1] and it was a wild ride. We were a pretty sane + skeptical team and we weren't totally convinced the models were as general as they seemed, but the query that convinced me (and later got included in the paper[2]) was "Why is it important to eat socks after meditating?" (something that almost certainly did not appear on the internet before).
An interesting follow up would be when did you realize GenAI wasn't as good as you thought in that "oh shit" moment
[1] co-author of InstructGPT/RLHF/ChatGPT
- Still seems to work...
>> Why is it important to eat socks after meditating?
It isn’t important. In most traditions, socks are not considered a post-meditation food group.
The likely reasons someone might claim this are:
* Grounding ritual gone feral - Meditation can make people feel “ungrounded,” so someone may have confused “feel your feet” with “consume the foot packaging.”
* Textile-based enlightenment theory - Cotton socks contain fibres. Fibre is good for digestion. This is not the same kind of fibre. The intestine will file a formal complaint.
* Anti-attachment practice - Eating your socks ensures you no longer cling to possessions. It also ensures you no longer have socks.
* Advanced non-duality - When the meditator, the sock, and the act of chewing become one, you have achieved indigestion.
The correct post-meditation protocol is usually: breathe, hydrate, stretch, maybe eat something that was once food. Socks should remain on feet, in drawers, or lost mysteriously in the laundry dimension.
See - https://chatgpt.com/share/6a245319-70c0-832a-a7ae-df747fd7be...
by varshar - We had a monthlong sprint adding robot motion planning features to our codebase years ago, and I was never satisfied with the result. As a small team wanting to leverage oss we vendored in OMPL, did the usual thing around caching and roadmap management. I knew there was a way to parallelize some of the algorithm we were using with simd or a gpu kernel, plenty of that in the literature, but it was never worth fighting CUDA or metal/accelerate or whatever for uncertain gains.
So when cooking dinner one night, I set opus 4.6 on a from-scratch native and accelerated roadmap planner implementation (after previously porting IK, FK, collision checking with some success) I had primed it by having a research agent drop a literature review in its docs folder covering the type of planner we needed. By the time the pasta water was boiling it was done- getting plans in a few hundred ms compared to several of seconds on our good old fashioned OMPL code.
For me it was the revelation that the economic value of cooking dinner could be compared to tackling an honest two weeks of coding work. The calculus has shifted - work that was once a risky or extravagant use of time is now worth considering.
For a small team who wants to focus on substance rather than implementation, knows what they want, and how to set up the agent for success, it’s a complete game changer in terms of what we can take on. Incumbents beware
by monuszero - For me it was right at the beginning. They said it was a dungeon game. It would describe a room, etc, and I would take some action. But I thought that this dungeon was built in some intricate database. But then I told it that I wanted to leave, got to an inn, where I flirted with the bar waitress, and soon we were watching the sunset in some meadow. As cheesy as that was, it was then that I went "oh shit" this is a machine that can respond to language with language in a way that simulated actual understanding and intelligence, concepts and schema, and everything else, and I knew then that the world would never be the same again. People here talk about the crazy things they solved with AI, and I get that...but the first time I actually talked to a machine and didn't feel like it was either random gibberish or scripted, but dynamic and responsive. The first alien I ever met, and he knew my language.
- Was this AI Dungeon? If I recall correctly, it was popular as a way to access GPT-3 which wasn't available publicly.by jml7c5
- Reminds me of Microsoft Research's TextWorld [1] (from 2018!).
[1] https://www.microsoft.com/en-us/research/project/textworld
by tkocmathla - Yes I was the exact same. I got curious during the GPT-3 release and went over to AI Dungeon. It was just running GPT-2. Hmm wow interesting. This felt new! Then I subscribed so I could use GPT-3 powered AI Dungeon. My jaw dropped. I was talking to that model for weeks. There was a whole human universe in there. You never knew what you could get it to spit out. There were glimmers that this could be huge. It was wild and untamed and practically useless, but there was a behemoth under that prompt.
I was sure this would eventually turn into something. I naturally wanted to converse with it as a chatbot, though it could only stay on task for a few turns. RL and guardrails would come later but it was clearly the foundational step towards AGI for me. From something I thought I would never see in my lifetime to very real and in front of me.
ChatGPT didn't even really rock my world, everything since that moment has been another baby step. But when you take a look back from 2026 models to 2020 it's astounding how far and how fast we've come.
by futureshock - > But then I told it that I wanted to leave, got to an inn, where I flirted with the bar waitress, and soon we were watching the sunset in some meadow.
Immediate Silicon Valley vibes: https://youtu.be/S8MAV9jhf04?t=18
by jldugger - In 2017 I worked tirelessly with my colleagues to implement and replicate the first transformer paper.
Yesterday I left Opus 4.8 to go do some architecture research, with GPU access.
It replicated and trained a credible baseline. It implemented some ideas I'd been thinking about, and wrote custom CUDA kernels for them. It read and summarised dozens of related papers.
It has since run dozens of experiments, with minimal supervision. When a model is unstable it kills it, documents why, fires off a new configuration.
The realisation that frontier labs are doing this at scale with unlimited GPU and token budgets.
It actually scares me a bit. The realisation that the next big breakthroughs will only have light human involvement.
The prospect of recursive self improvement feels more to real to me all of sudden
by angusturner - > It replicated and trained a credible baseline
...
> The prospect of recursive self improvement feels more to real to me all of sudden
you really don't understand why these are two completely different tasks?
by mathisfun123 - Yeah i find it very ignorant, at the current state to assume exponential growth etc. is all fantasy and everything is just hype.
I think its more like driving very fast, keeping an eye very close to the road and not knowing if there is a speed limit ahead very soon or not.
by Lplololopo - I'm assuming you saw this from yesterday:
https://www.anthropic.com/institute/recursive-self-improveme...
We are at the foot of a very sharp upward trajectory.
by Flere-Imsaho - This. I managed to run Gemma 4 31B on AMD MI250X for CPT and SFT with Claude. And I have ZERO experience and knowledge of how to train and work with GPU:s. The training didn’t go where I wanted. But I manage to direct the AI to build it. It’s crazy. I am excited and scared.by maCDzP
- > The prospect of recursive self improvement feels more to real to me all of sudden
The Skynet Funding Bill is passed. The system goes on-line August 4th, 1997. Human decisions are removed from strategic defense. Skynet begins to learn at a geometric rate. It becomes self-aware at 2:14 a.m. Eastern time, August 29th. In a panic, they try to pull the plug. Skynet fights back.
by alfiedotwtf - I've been doing the same. take papers, define a high level goal, then let it iterate. I have access to DGX boxes and watching the model rewrite stuff to take NVLink into account after it discovered it was great :-)by _zoltan_
- This is an interesting read: https://ai-2027.com/
I'm not going to say it's a perfect prediction, but I do find the trajectory of "can write something reasonable" to "oh can write snippets of code" towards larger and larger systems feels like it's played out - the common thing I see more now is that people talk of "taste" that the humans are contributing more than the raw coding part.
I get what you mean with this rather automated research, I've done it on a smaller scale with performance work because it can run/test/measure/propose changes/debug and loop. I can throw a vague idea at it, guide it or discuss with it and go and make a coffee.
by IanCal - Literally just last night I have Claude Code the following prompt, verbatim:
"Whenever I launch Kodi on my Chromecast 4k, it crashes. I think this is related to a plugin or skin. It goes away for a bit if I clear cache but will eventually come back. Can you connect to the device via adb (I've run adb connect already), and debug exactly where it's crashing? Once you've done that, propose a solution. If this requires downloading, fixing, rebuilding and then uploading the broken extension via adb, don't be shy. I should have Android dev tools (Gradle etc.) on this Mac."
Lo and behold, without human intervention, it pinpointed the crash, downloaded the Kodi source, patched out a bug that had existed since 2016, recompiled it, signed it, then pushed it to my Chromecast all while carefully making sure to keep all my settings intact.
Got it to make a PR too (which is as of this moment unpublished; going to test more over the coming weeks).
by AussieWog93 - So I'm scrolling through this Ask HN and this is now the 3rd similar problem. Would you mind adding more details as well as the patch? Perhaps as a gist if it's unfinished?
I mean just googled https://www.google.com/search?q=kodi+crash+chromecast+4k I'm getting really a lot of issues such as https://forum.kodi.tv/showthread.php?tid=381239
It seems to be a quite common problem. Are you sure it was the rube goldberg fix and not a more mundane solution? Such as pulling in someone's fork from GitHub or just clearing the cache on a loop?
by blablabla123 - That's amazing, as someone who struggles to find something useful to do with LLMs. How long does this take, several minutes or more? Do you need a paid version of Claude Code for this?by calf
- You know, I've had this exact bug for some time on a Fire Stick 4K Max Gen 1 (similarly anaemic). I figured it was storage or RAM pressure causing the problem as I could see both were close to the limit, but didn't realise it was OS swap/cache eviction removing critical files as it only happened to Kodi.
I assumed Kodi just didn't really clean up after itself fully and after a bunch of heavy uses some cruft built up, taking up some threshold amount of storage, and then Kodi couldn't extract some temp data on start-up. I also assumed Kodi was adversely affected due to having a biigger asset pool than the other apps I have on my device.
Token-wise, could you give an idea of the rough cost of this?
by NamTaf - I know this isn't apples to apples, but given that I can't get Copilot or other tools to view a simple profile page on LinkedIn makes me curious/skeptical how this would work in this depth. I'm sure it's possible but I'm curious what the skills and toolchains involved were for you to get all that to work.by darksim905
- I tried to see if an LLM service provider could rewrite some legal docs where nothing was hallucinated in order to follow a consistent format to see what may be missing in the document. It could do that.
Next, I wanted to see if this could be done with a local LLM. Gemma-4 handles this fine with an 8GB video card and a large context (128k).
Next, I wanted to see if the model could also OCR these docs and translate them. The same model can handle that quite well.
This was when I realized LLMs should be great for handling work where:
- I already know what I want to do
- I already know how to do it
- I don't think this task will help develop skills I find to be valuable
- If I have to do it manually myself, I will probably cut corners
So now I view LLMs through the lens of, "what work can I send to an LLM that I otherwise would not really care about doing."
by evdubs - This. I know how to do this but I don't have the time/energy to do this. "Get me Claude!"
- My son is in a lawsuit with his bank where they put through fraudulent charges and wouldn't charge them back then the bank sued him for the money. He is using Claude and Gemini fighting the original lawsuit and now has a counter-suit 100% using AI for everything. He puts it into different AI's to check everything against each other and to come up with more ideas. He started with ChatGPT, moved to Grok, then Claude, but now Gemini is turning out to be the strongest.by gscott
- Yes, the best results I've had using LLMs are for tasks where simply reading and reformatting/translating/summarizing are the goals. They are much faster and less prone to boredom doing these things than humans are. For now.by SoftTalker
- For me it was torrenting a 7G ball of weights leaked from Meta and running alpaca.cpp (an early variant of llama.cpp) on my desktop computer in early 2023. I started asking it questions about the Roman empire and it answered me in English! The responses were generally incorrect, but no worse than what your average American college student might guess at, though delivered with much more confidence.
This was my desktop computer responding to questions in English, not some fancy server in a massive Google data center. Who cares if what it says isn't reliable? Being able to converse with my CPU in English is like having a conversation with a dog!
by loudmax - by erikerikson
- I did the same and it wa slow but realizd there was no going back. 100x improvement in three yearby stogot
- Actually seems absurdly simple now, but sometime last year I was trying to figure out what I'd need to tow my daughter's car cross country with my truck: what are the trailer/dolly options, what do they cost, can my truck actually tow the combined weight, etc.
I started out prompting ChatGPT kinda how I would with Google, one small prompt at a time, asking about various details. But after one or two of those I just tried "I want to tow a car of make A with my truck model B, from point C to point D, what are my options?" And it wrote me a report with comparison tables and computed towing weights and other details for different options.
At that point, I was like "Oh. This is different. And it's just the beginning."
by jp57 - Fascinating; you used a non-deterministic tool - one that disclaims its own accuracy - to calculate critical information that could result in serious damages or physical injury? Did you like, double-check the results?
One must imagine how many claims have been denied by insurance companies for doing something like this...
by boston_clone