As an engineer, I love the idea of flipping the problem on its head. Don't get rid of the engineers; get rid of the CEO. That's brilliant.
But if we are seeing problems with agents replacing knowledge workers at the coding layer, why assume the executive layer is uniquely immune?
There was an interesting article in the Harvard Business Review about this exact problem. Researchers found that when LLMs are asked for strategic business advice, they tend to converge toward what is popular and trendy rather than necessarily what is best for the particular situation. They called this "trendslop," and described the response as,
"On strategy, LLMs might be more akin to a freshly minted MBA or junior consultant, parroting what's popular rather than what's right for a particular situation."
Counterpoint: most human CEO's are doing trendslop as well. At least in the worlds I live in (US technology industry, so either megacap public companies or SV-funded startups). It often seems like all of the VCs, CEOs, and large investors in public software companies are all on the same group-chats and sharing the same ideas with each other, driving lots of this trendslop.
From observation of having worked at 7-8 companies as a software dev, it sure seems to me that most CEO decisions could be replaced by AI easier than software dev work can.
I don’t understand why people are calling these transformers models non-deterministic? Are you referring to the temperature parameter? I haven’t played with transformer internals in a while but my understanding is that if the temperature is fixed at a value where the top logit is always picked, then because they weights are fixed, the exact same input should produce the exact same output. Am I missing something?
> if the temperature is fixed at a value where the top logit is always picked, then because they weights are fixed, the exact same input should produce the exact same output. Am I missing something?
Yes.
Your input is part of a batch, and you don't know where in the batch it is. By default batches are not invariant and VLLM only supports invariance at all on some Huwaei Ascend hardware.
By non-deterministic I think people really mean "chaotic" in the chaos theory sense. Small perturbations in the input lead to wild and unpredictable changes in the output. Even with temperature parameters a fixed PRNG seed could mean an LLM was just chaotic and not technically non-deterministic.
But more literally while LLMs are in theory deterministic (though perhaps not inference providers implementations if there's anything like a race condition affecting how things are rounded when added together) - we use the LLMs in harnesses that aren't. There are very likely races in the terminal outputs, dates both intentionally put in the context and accidentally leaked to the context, things like that.
Ok. I see. I guess people are not referring to the raw models themselves when they say non-deterministic, but are also including the harness used in conjunction with the model. Then, in that case, for the exact same input you could get a non-deterministic output. But the model itself and all the mathematical machinery around the model is still very much deterministic.
I guess if we really needed to, we could construct a deterministic agent harness. But in most use cases we probably want some chaotic behavior to increase our chances of stumbling on the desired results.
I think people are wrapping that across the English language. In English, these two tasks are exactly the same:
"Would you hand me that item?"
"Please hand me that item"
But when posed to the LLM, they generate different outputs. One character difference in the prompt might be a whole different output. People who aren't programmers mostly don't know that there's any difference. They asked for the same thing, it knows what they want in both cases...but different results.
I'm not sure that's true. Sure, in the end I might hand them the item, but my thoughts about what they said will be different. I think you have to consider my thoughts "output" for this comparison to be valid.
Not to be too pedantic, but these requests would not be exactly the same.
There is a bit of indexicality in "Would you hand me that item ?"
that might cause it to be interpreted as an actual question rather than a request, and might elicit different responses:
- maybe _I_ would not hand this to you (I'm busy right now), but the person next to me whose hands are free would, so I'd nod to them. However, if you had said "Please hand me that item" I'd put down what I was doing to comply.
- maybe I would not hand _this_ to you (it's not the right tool IMO), but I'd suggest another option. However, if you had said "Please hand me that item" I'd put my doubts aside to comply.
- maybe I would not hand this to _you_ (you're not the one who should be handling it), but I'd do the thing myself or hand it to a more qualified member of the group. However, if you had said "Please hand me that item" I'd trust you enough to comply.
I think this distinction is relevant in that I've found people to sometimes have difficulties understanding how similar LLM prompting is to giving instructions to human colleagues.
I've had a collaborator who though very highly of his own prompting skills (while his prompts were very ambiguous and of the "make no mistakes, erase everything & correct yourself if you find one" variety) and blamed the models for not being "smart enough", and it was very noticeable that his management style for the juniors on his team was similarly unproductive.
Depending on my mental state, status with the person and many other factors each of them may trigger both many different internal thoughts, looks, body expressions and even outcomes.
As the above two comments mentioned this is not true in practice due to batch effects (you can read about some interesting work published by Thinking Machines on this), as well as calculation drift that happens across computations esp. now with inference optimization becoming common.
I remember those days. They released smaller versions of the model that were interesting but pretty useless. The wild part is, about 5 years later I was able to train the same model from scratch, using a couple of gpus in my home office, following Kaparthy’s online lectures, for a fraction of the cost. Long Live Open Source!
I agree. I like using Antigravity for some of my frontend work, and I find it does a better job than Claude Code - Opus 4.6. I’ve also found the Gemini Flash models to be good at legal defense research—I use them to help New Yorkers fight parking tickets (https://nyceasyparking.com). That said, the Claude models are still amazing at agentic work.
I’m working on yet another cloud based coding agent https://seniordev.io/ that connects to an existing GitHub repo, spins up a feature branch, commits incremental changes, and opens a PR. You can jump into an embedded VS Code server to review and tweak the code before merging—no local setup needed. Any feedback is greatly appreciated Thanks!
Great work! I love your videos; they've taught me so much. Any plans for a Mixture of Experts (MoE) video? My understanding is that starting from GPT4 most advance models use MoE to some extent. For example, can I take the model from your GPT2 video and just change the feed forward layer to an MoE layer like the one found here (1)? I guess I can just try it myself but I enjoy the expert guidance you provide in your videos. Please don't stop! great content!
The 555 timer is considered one of the most successful chip designs in electronics. It has a simple, well understood, design. The timing on the 555 can be configured to generate a periodic signal using just a few resistors and capacitors. The 555 is often used as a timer in electronic circuits. The Linux scheduler is usually configured to respond to interrupts generated by a cpu. To replace those cpu interrupts by a simple 555 timer is impressive due to the unconventional nature of the setup.
Also, the 555 is an ancient chip, that pretty much appeared as the basic chip (along with it's friend, the 741 general purpose op-amp), being the two of them the protagonists of pretty much 70% of the hobby circuits in the 80s.
Sure thing! One of the things that helps me sometimes is re-arranging things visually.
A real divider:
Input o----vvvvv----o-------------o Output
|
|
Z
Z
Z
|
|
o
Ground
The bad one:
Input o-------------o----vvvvv----o Output
|
|
Z
Z
Z
|
|
o
Ground
Looking at it this way, it might be a bit easier to see that in the good circuit, both resistors can be in a Kirchoff voltage loop that includes the input, while in the bad one they can't.
(Edit: it appears I don't understand linebreaks on HN, so things are gonna appear garbled until I fix it.)
Hi apologies for any confusion. What i meant to say is that OpenWrt can be run on the Raspberry Pi. I did this because I had a Raspberry Pi laying around and did not want to invest in additional hardware just for experimenting with OpenWrt within my internal network. You can also run OpenWrt in a virtual environment like QEMU -- this is useful if you're aiming to connect multiple VMs to OpenWrt, effectively setting up a virtual network with OpenWrt at its core. Please see links below for more info:
It could be that the corpus ChatGPT was trained on is full of ‘confidently wrong’ answers from these ‘experts’. One solution could be to train these LLM on a higher quality corpus from real experts instead of random text from the internet. But would that just bring us back to the days of expert systems?
That will not solve the problem, because when GPT doesn't have the answer, it will make one up by copying the structure of correct answers but without any substance.
For instance, let's say your LLM has never been told how many legs a snake has, it knows however that a snake is a reptile and that most reptiles have four legs. It will then confidently tell you "a snake has four legs", because it mirrors sentences like "a lizard has four legs" and "a crocodile has four legs" from its training set.
I don't think this is necessarily the case anymore. The bing implementation of chatgpt has a toggle for how cautious it should be with getting things wrong. I was working on a very niche issue today and asked it what a certain pin was designated for on a control board I am working on. I believe it is actually undocumented and wanted to see what chatgpt would say. And it actually said it didn't know and gave some tips on how I might figure it out. I suppose it is possible that it synthesized that answer from some previous q&a somewhere, but i couldn't find any mention of it online except for in the documentation.
But if we are seeing problems with agents replacing knowledge workers at the coding layer, why assume the executive layer is uniquely immune?
There was an interesting article in the Harvard Business Review about this exact problem. Researchers found that when LLMs are asked for strategic business advice, they tend to converge toward what is popular and trendy rather than necessarily what is best for the particular situation. They called this "trendslop," and described the response as,
"On strategy, LLMs might be more akin to a freshly minted MBA or junior consultant, parroting what's popular rather than what's right for a particular situation."
I really hope this idea works but just thought I'd share some of the research out there. https://archive.is/OPObK/again?url=https://hbr.org/2026/03/r...