hasteg

I did RL and evals at Amazon for a bit on Nova Forge (laid off in July! :D). I would have to push several thousand lines of code a day in new evaluations that our applied scientists requested since new ones are always being made along with having to create custom ones for whatever new model they were working on. Also, in RL, adding new environments for the agents to actually complete activates was also a shit ton of lines of code. So if I had to guess, just adding new stuff around model training and testing alone accounts for a ton of these changes.

Made something similar a few weeks ago with GPT6 when it launched to test it out, however for just US lighthouses: https://jmoul.io/lighthouse-map/?v=20260909-3

Although this one look a lot better. cool! like how the beams look like they're ray traced.

I feel like this is a cool application of using ai - before something like this would be unfeasible to make due to the amount of lighthouses. Parsing through the amount of data to create a description of every lighthouse would take months+.

This is likely being solved with stuff like watermarking which Anthropic just added to Claude... I'd imagine they are testing new data and verifying that training data (or at least the data which do NOT want to be AI generated, i.e actual human text) is not generated, at least with claude.

Right… unfortunately my 9800x3d and 5090 setup can’t process a 15 year old desktop app transition in under a second even though they have a coding god making all their decisions hahah

Alright, seemed like it was a large jump the OP was making. Just was interested in knowing the reasoning for knowing why they thought that.

Well I mean, based on what I said, if we truly did have “AGI” then the “AGI” would be able to argue it’s 100% correct point in court according to copyright law, no human intervention. Yes there is lawyers and judges who are making calls in court but allegedly it would be able to construct a concrete argument right? That would hold up in court. Without any buts or ands. Not disagreeing with you, I agree as of now it wouldn’t hold up. But a true “AGI” would be able to pander to anyone better than any PERSON could.

Is this true? Any sources? Been buying a lot of gold over the past year or so.

My favorite part about the desktop app is the 1 second delay between when I click a playlist and when it actually loads.

That’s a good point… you would think AGI would be better at being a copyright lawyer than any human would be and as such would be able to distinguish whether something is “copyright infringement” or not… so the system prompt should just include “make judgement calls on reproducing copyrighted material under the full scope of the legal framework in place” or something.

I've been running 27B a lot, I am honestly shocked at how well it performs. It's mind blowing how well the small qwen models (and particular, the 3.8 model) runs locally. It can create some really impressive toy coding projects. I haven't really done too much integrating into my actual workflows because I pay $100 for Claude Max, but I can see it being pretty helpful in that.

As a 5090 owner and local model enthusiast, I was hoping it would be 35B A3B so I could run it myself =(.

I live right next to a micro center and remember when they started offering that deal... still so pissed at myself for not buying one. I ended up just buying a raspberry Pi for what I was doing, but seeing as where the prices are now, I messed that up a bit. Also my worst sin was not buying 64GB of DDR5 when I was doing my computer upgrades back in August last year.

This is pretty cool. Implemented something similar myself (a really small language model with ~10M params) just to teach myself the ML behind the LLMs. Did not implement it in C obviously, just use PyTorch, but it's interesting to go through the c file and see how he has implemented stuff I took for granted in Python in C.

Anyway, I just tested this out myself on my AMD Ryzen 9 9800x3d. I got 7647173 tok/sec using karpathy's Shakespeare dataset https://raw.githubusercontent.com/karpathy/char-rnn/master/d.... Going to play around with it and see if I can get a CUDA kernal built to see what it could do on a 5090. Claude estimates with napkin math that we could get around 2B tok/s

At this point I believe that when people do not cite using AI to generate their comment it's from a position of malice... like you're intentionally trying to position yourself as capable in whatever subject you're talking on but realistically you are not given the reliance on an LLM. Not saying anything about OP but just a general observation/opinion I've been building recently. Seems more common than ever to larp using AI.

It's so obvious when a book is generated by an LLM that I would blame the person buying the book moreso than I would the person who generated it for it. You can tell in a few sentences if something is AI generated (and by proxy - if it's worth reading or not).

I think using LLMs to research is completely fine, as long as authors would take the information and synthesize their own opinions based on the findings. I read books/articles/posts/comments to hear other PEOPLE's opinions, not LLM generated opinions.

What country/region of country (if US) do you live in? Everyone I know absolutely hates it and hates the direction things are going. I'm pulling from a wide sample here too -- I have a lot of friends who work blue collar jobs that hate it just as much as my tech worker friends. However I will say I'm in my late 20s and it seems like my age group and younger are more anti-AI than any other age bracket.

I agree that LLMs can make good points and defend their arguments, but if I'm talking (or debating) with someone online and it's obvious that they are just responding with LLM output, it immediately discredits them in my mind and is a pretty clear indicator that the actual person doesn't have a good understanding of what they're talking about... and I can go argue with an LLM myself. If I'm on reddit for example and someone starts obviously posting verbatim LLM output, there's no point to really continue to engage or read. I'm not saying it's wrong to use AI in these situations, but IMO, I'd at least expect the person using it to at least read the output and then use it to synthesize their own response in their own words. Literally copy and pasting a LLM output verbatim just signals to me that the person using it doesn't have the capability themselves take information and use it to formulate an opinion/idea/argument/whatever on the given topic. I'm interested in talking to the person, not an LLM.... I can go prompt the LLM about whatever topic myself.

Lol, love this post. Every single day when I take the elevators at the office I try and figure out what the algorithm is behind it. Been too lazy to just research it. Now I know!

I agree that using AI as a "tutor" to help teach yourself things is very useful and works great, I've been doing that myself quite a bit. How do you stop students from being able to just use it as a second brain? I assume there's some sort of guardrails you have so it won't just give the answer? How do you ensure that there's an adequate amount of "struggle" that the kids are going through before just explaining the answer? I feel that that would be the challenging piece of the puzzle.

This is also how I pretty much filter LLM generated text in my head.

Using AI to actually learn is indeed very helpful I've found... I just don't think most people use it for that. I recently wanted to build and train my own (albeit, small and simple 10M param model) and used AI extensively to explain concepts for me, explain lines of code, and generate in depth visuals I can use to get a visual intuitive understanding of what is happening under the hood. I think people who have a natural curiosity to understand the why and how of things benefit immensely from it, but I do admit it is very easy to just offload all thinking to it instead of asking it why something is happening. I could have just asked it to implement the entire model using PyTorch and just ran the CLI command to begin training on some auto-generated dataset, but intentionally struggled through it. Actually learning from it requires intentionality for sure.

I just recently watched some (not all) of this video "coding a machine learning library in c from scratch" and seems like he's going through a similar process in this blog as this video. I would recommend watching the video to get an idea of what the fundamentals of a ML library look like. From someone who has recently been getting interested in actually writing ML code and trying to make sense of it myself (from the perspective of just a typical backend engineer) it was very interesting to see. Previously my experience with ML libs (PyTorch specific) was writing my own Mini-GPT and training it on a small dataset using my own GPU (5090). Cool to see the behind the scenes and took away some o the handwaveyness... https://www.youtube.com/watch?v=hL_n_GljC0I

Yeah, clear example of survivorship bias... not to sound negative, but if Facebook had become nothing, odds are Boz would be nothing more than another engineer. It's all too common a tale to hear people spout their nonsense and think they got to where they are because of some inherent talent or skill, but at the end of the day, getting to somewhere like Boz these days is just luck. Does anyone really think you could start as an L4 at something like Amazon these days and end up as a CTO? Nope.

Same here, I hate it. Instead of just reading I find myself also dedicating brain power trying to decide if it's worth reading or not based on the first few sentences... if someone can't put in the effort to actually write something themselves I absolutely do not want to put in the effort to read it.

LinkedIn is basically unusable at this point. I actually did used to use it a fair amount before but I've since deleted it and just use email notifications to see any notifications from recruiters.

What I don't get is how these people don't feel shame in their super obvious blatant use of LLMs for everything, even responding to posts. Maybe it's just me but when I'm attaching things to my name like that, I would absolutely not want everything to be obviously slop shit. Do they think people can't tell or something? I know at least every technical person I know can immediately tell (most of the time) when writing is LLM generated.

I mean, this is perhaps the most vibecoded website possible so... deff AI.

Can you elaborate on that more? Why just a filler character?

Really interesting paper, thanks for the share.

The point their making in that paper reminds me of this paper some people shared around work earlier this year, https://arxiv.org/pdf/2512.14982 (Prompt Repetition Improves Non-Reasoning LLMs)... I wonder how OPs question would fare (or the questions presented in the paper you posted) given double repetition.

Haha that's a great intro to it. I remember very clearly I was obsessed with minecraft griefing/team Avolition at the time and found it to ask if someone could make me a hacked minecraft client for my first thread (yikes). Wasn't a great intro to the community for me but ended up staying after that. Also was very significant to my computing career, glad I was on it. I think that era of the internet was very special and I miss it dearly. Also seems like back in 2015 or something someone somehow got into my account and Omni deleted it. Sad.