If you have a decent understanding of how LLMs work (you put in basically every piece of text you can find, get a statistical machine that models text really well, then use contractors to train it to model text in conversational form), then you probably don't need to consume a big diet of ongoing output from PR people, bloggers, thought leaders, and internet rationalists. That seems likely to get you going down some millenarian path that's not helpful.
Despite the feeling that it's a fast-moving field, most of the differences in actual models over the last years are in degree and not kind, and the majority of ongoing work is in tooling and integrations, which you can probably keep up with as it seems useful for your work. Remembering that it's a model of text and is ungrounded goes a long way to discerning what kinds of work it's useful for (where verification of output is either straightforward or unnecessary), and what kinds of work it's not useful for.
crystal_revenge · 1h ago
I strongly agree with this sentiment and found the blog's list of "high signal" to be more a list of "self-promoting" (some good people who I've interacted with a fair bit on there, but that list is more 'buzz' than insight).
I also have not experienced the post's claim that: "Generative AI has been the fastest moving technology I have seen in my lifetime." I can't speak for the author, but I've been in this field from when "SVMs are the new hotness and neural networks are a joke!" to the entire explosion of deep learning, and insane number of DL frameworks around the 20-teens, all within a decade (remember implementing restricted Boltzmann machines and pre-training?). Similarly I saw "don't use JS for anything other than enhancing the UX" to single page webapps being the standard in the same timeframe.
Unless someone's aim is to be on that list of "High signal" people, it's far better to just keep your head down until you actually need these solutions. As an example, I left webdev work around the time of backbone.js, one of the first attempts at front end MVC for single pages apps. Then the great React/Angular wars began, and I just ignored it. A decade later I was working with a webdev team and learned React in a few days, very glad I did not stress about "keeping up" during the period of non-stop changing. Another example is just 5 years ago everyone was trying to learn how to implement LSTMs from scratch... only to have that model essentially become obsolete with the rise of transformers.
Multiple times over my career I've learned lesson that moving fast is another way of saying immature. One would find more success learning about the GLM (or god forbid understanding to identify survival analysis problems) and all of it's still under appreciated uses for day-to-day problem solving (old does not imply obsolete) than learning the "prompt hack of the week".
thorum · 1h ago
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.
PaulHoule · 1h ago
Rewritten in response to quality complaint.
A lot of people are feeling HN is saturated with AI posts whether it is how MCP is like USB-C (repeated so much you know it is NPCs) or how outraged people are that their sh1t fanfics are being hoovered up to train AI.
This piece is not “news”, it’s a summary which is tepid at best, I wish people had some better judgement about what they vote up.
Fraterkes · 1h ago
Just a heads up: you should try to get better at writing.
PaulHoule · 54m ago
Thanks for the tip!
alphazard · 59m ago
When explaining LLMs to people, often the high level architecture is what they find the most interesting. Not the transformer, but the token by token prediction strategy (autoregression), and not always choosing the most likely token, but a token proportional to its likelihood.
The minutiae of how next token prediction works is rarely appreciated by lay people. They don't care about dot products, or embeddings, or any of it. There's basically no advantage to explaining how that part works since most people won't understand, retain, or appreciate it.
helloplanets · 32m ago
It's not a model of text, though. It's a model of multiple types of data. Pretty much all modern models are multimodal.
panarchy · 34m ago
AI research was so interesting pre-transformers it was starting to get a bit wild around GPT2 IIRC but now the signal to noise is so low with every internet sensationalist and dumb MBA jumping on the bandwagon.
qsort · 1h ago
I agree, but with the caveat that it's probably a bad time to fall asleep at the wheel. I'm very much a "nothing ever happens" kind of guy, but I see a lot of people who aren't taking the time to actually understand how LLMs work, and I think that's a huge mistake.
Last week I showed some colleagues how to do some basic things with Claude Code and they were like "wow, I didn't even know this existed". Bro, what are you even doing.
There is definitely a lot of hype and the lunatics on Linkedin are having a blast, but to put it mildly I don't think it's a bad investment to experiment a bit with what's possible with the SOTA.
crystal_revenge · 1h ago
> I see a lot of people who aren't taking the time to actually understand how LLMs work
The trouble is that the advice in the post will have very little impact on "understanding how LLMs work". The number of people who talk about LLMs daily but have never run an LLM local, and certainly never "opened it up to mess around" is very large.
A fun weekend exercise that anyone can do is to implement speculative decoding[0] using local LLMs. You'll learn a lot more about how LLMs work than reading every blog/twitter stream mentioned there.
I mean I didn’t find out about Claude code until like a week ago and it hasn’t materially changed my work, or even how I interact with LLMs. I still basically copy paste into claude on web most of the time.
It is ridiculously cool, but I think anybody developer who is out of the loop could easily get back into the loop at any moment without having to stay caught up most of the time.
qsort · 1h ago
I'm not talking about tools in particular, I completely agree that they're basically fungible, and for "serious" stuff it's probably still better to use the web interface directly as you have more control over the context.
The problem I see is that a lot of people are grossly misaligned with the state of the art, and it does take a bit of experimentation to understand how to work with an LLM. Even basic stuff like how to work with context isn't immediately obvious.
Fraterkes · 1h ago
I don’t think you’re wrong, but if it takes someone a month (at most) to get up to speed with these tools, I don’t think thats much of an argument for closely keeping up with them (until you need to know them to keep your job or smt) especially because everything is changing every few months. There is arguably no technology that needs you to “keep up” with it less
rglover · 24m ago
You don't need to "keep up," you just need to loosely pay attention to identify things/features that will make you more productive, test them out, and keep what actually works (not what some influencer claims to work on social media). In fact, I feel far more confident in my understanding by listening to researchers who dismiss the wild claims about AI's potential—not hype it blindly [1].
There's far too much noise, churn, and indecision at this stage to get any productive value out of riding the bleeding edge.
If it's actually revolutionary, you'll hear about it on HN.
The article doesn't actually explain the "why", which severely undercuts the existence of the "how" list.
It's fine to go do other things with your precious time instead.
pizzathyme · 1h ago
Exactly. I have been questioning the need to "keep up". As the best AI innovations arrive, the ones that are useful will eventually make their way to the mainstream (and me). I didn't "keep up" with the nascent development of spreadsheets or Google Docs. Once they were widely used, I adopted them just fine, and haven't missed out on anything.
Unless you are building an AI startup furiously searching for PMF before your runway expires, I don't see the urgency.
gowld · 48m ago
Someone who keeps up on AI and finds productivity gains will outcompete someone who doesn't, even for activities that aren't develoing new AI weren't AI-based before.
What is an "AI startup"? If you add a chatbot do your product, are you an "AI startup"? Does "startup" require having no moat? Can you be an established stable business that loses everything to a startup that leapfrogs your tech, AltaVista?
pamelafox · 2h ago
Great list! I also subscribe to Gergeley Orosz' "Pragmatic Engineer" which covers many AI topics now, and to Gary Marcus' substack, which tackles topics more from an LLM skeptic perspective.
Plus I subscribe to updates from Python packages like Langchain and PydanticAI to see what they're up to, since that's usually reflective of what's happening in broader industry.
I'm not on X anymore so I can't follow many of the folks directly, but some of them (like Simon Willison) also post on BlueSky and Mastodon, fortunately. Some folks like Sebastian Raschka and Chip Huyen also post on LinkedIn. Kind of all over, but eventually, I see a good amount of what's happening.
jameslk · 2h ago
Maybe I’ve been missing some important stuff, but it seems the most relevant and important updates eventually just bubble up to the front page of HN or get mentioned in the comments
SoftTalker · 2h ago
Trying to keep up is like jumping on a 90mph treadmill. I have decided to opt out. I think AI (and currently LLMs) is more than a fad and not going away but it's in a huge state of churn right now. I'm not investing a ton of time into anything until I have to. In another few years the landscape will hopefully be more clear. Or not, but at least I won't have spent a lot of time on stuff that has quickly become irrelevant.
I'm currently not using AI or LLMs in any of my day-to-day work.
bluefirebrand · 1h ago
I agree pretty strongly
Yeah it's not a fad, but I think it's really not as useful to me right now as the hype seems to suggest
I'm going to keep an eye on developments, but I'm not using it for my day to day either. I'm just not seeing the same value other people seem to be seeing right now and I'm not going to exhaust myself trying to keep up
One day Claude is the best. Next is Cursor. People are switching tools every two weeks trying to keep up
Not for me thanks. Especially not with how inconsistent the quality and output still are
chasd00 · 47m ago
i'm in the same boat, i did the zero-to-hero series karpathy put out to get a basic understanding of what we're dealing with. At work, I helped put together a basic rag setup last year that's in production so feel ok with implementation. Finally, i have a little personal pet project that i'm going to feel out with claude code over Christmas to get more hands on. That's about it all i'm going to do to "keep up". I'll let the fighters fight it out and then just use whatever wins.
edit: i use claude and chatgpt day to day to help with simple things like regex, a replacement for google search in same cases, and self contained classes, functions, and other small discreet blocks of code.
HellDunkel · 1h ago
This. When has early adoptation paid off lately? Remember prompt engineering?
Disposal8433 · 1h ago
I remember the same FOMO that happened with "Google searching." People wrote and sold books about how to search properly on Google. The same with AI will happen: either it will flop, or we'll be able to learn all the needed skills in a few days. I don't worry about it.
Mars008 · 41m ago
For some of us this time has come, i.e. it's already useful and worth learning and spending on prime accounts.
gowld · 47m ago
I've banked professional success on my superior "Google Searching" ability (compared to my peers) for 20 years. A little but of thinking goes a long way, even when using tools designed to replace thought.
bzmrgonz · 1h ago
give me whisper real-time transcriber and I'll be happy!!
roboror · 1h ago
Yeah these companies have made unbelievable investments, keeping their products a secret is antithetical to their existence.
mettamage · 2h ago
This is my attitude for all tech surrounding IT
codebolt · 1h ago
Just subscribing to OpenAI, Anthropic and Google on YouTube is pretty helpful. They post demos of all major new feature releases that are good to skim through to get a sense of where the frontier is moving (just take all claims about capabilities with a grain of salt).
I've also got some gems from Microsofts Build talks, specifically whenever Scott Hanselman and Mark Russinovich get together, e.g.: https://www.youtube.com/watch?v=KIFDVOXMNDc
twapi · 2h ago
Simon Willison's weblog is almost enough, highest S/N
Indeed just get out of tech and make a new living! Tech jobs are declining and will continue then fall off a cliff with one doing the job ten use to. Followed by other white collar and blue collar (AMazons warehouse robots) jobs.
Happily canceled my GPT Plus this week; personally not gonna feed that beast any longer! As well it can not generate maps (create road trip travel maps showing distance between locations to share with friends, a creek tubing map route & etc) at all like Gemini can for free.
astrange · 1h ago
> Tech jobs are declining and will continue then fall off a cliff with one doing the job ten use to.
This would increase employment ceteris paribus. That's like saying inventing new programming languages is bad because they're too productive.
This might be true, but if it is, the one "doing the job ten use[d] to" would not actually being doing the same kind of work at all, and so therefore might not be the same people or even same kind of people. Even if we do Jevons ourselves out of this situation, it might still spell employment disaster for mid level coders, while increasing employment for product managers.
astrange · 15m ago
Indeed that's the case, but it's sort of what software engineers signed up for anyway. I don't know if people are expecting to go their whole careers without learning a new stack / language. Maybe they are?
dingnuts · 1h ago
That's right, there are no carpenters or lumberjacks anymore because power tools were invented
kybernetikos · 1h ago
There are lots of careers that used to be widespread but are now extremely niche because of technological change. For example, most careers relating to horses.
paul7986 · 16m ago
Exactly hence why i said its probably wise to not pursue tech unless your an AI researcher PhD level already, striving for such in school studies and or leveling up your skills in AI.
paul7986 · 28m ago
So say you are a startup are you going to now hire a designer or use WAY Less Expensive & quicker AI to design logos, website, an app, etc? Print design.. all types of design it can do.
So You and all other people like to save money are going to continue spend the same thousands on such a resource when AI can do what they do in a few minutes or more for WAY LESS? UX Design was supposedly a growing field ... not at anymore! Definitely one can do the same thing in that field that 10 did.
Further, future mobile AI devices will pull the information and put it all on the lock screen of your AI device visualizing the data in a fun new way. Technology that makes things simpler and more magical get adopted yet visits to websites will significantly decline.
For federal workers who have lost their jobs they are feeling this pain competing for jobs against each other and now AI. It will only get worse for designers because it's now cheaper and faster to use AI to design logos, sites, apps to even including do vibe coding for the front end development to possibly the backend but that's not my specialty yet no doubt I vibe coded front-ends.
ninetyninenine · 1h ago
Obviously AI is different. While LLMs are more of a power tool now, the future trendline points towards something that (possibly) replaces us. That's the entire concern right? I mean everyone knows this.
Is this not obvious?
Why do people hide behind this ridiculous analogy: "That's right, there are no carpenters or lumberjacks anymore because power tools were invented"
???
I mean sure the analogy is catchy and makes surface level sense, but can your brain do some analysis outside the context of an analogy??? It makes no sense that all of AI can be completely characterized by an analogy that isn't even accurate yet people delusionally just regurgitate the analogy most fitting with the fantasy reality they prefer.
lurking_swe · 42m ago
sure. But why is digging one’s head in the sand a good strategy? To be clear i’m not advocating for trying to keep up all the time. You gotta live life too. But “ejecting” completely is dumb.
Are you saying that when the sewing machine was invented, it would be in the employees interest to not learn how to use it? Or when the computer was invented, it’s not in the employees interest to learn how to use it?
Even if you are a software engineer and are fired / laid off / pushed out of the industry because of AI, knowing how to use AI, its risks, etc is still helpful. It’s a skill you can hopefully use in your next career, whatever you pivot to. Unless you pivot to manual labor.
Thinking otherwise is shortsighted.
muglug · 2h ago
Just follow the first bullet point (read simonw's blog) and you'll probably be fine.
dvfjsdhgfv · 2h ago
Too lazy for that, a Fireship accompanying the morning coffee must suffice.
krat0sprakhar · 2h ago
+1 - super high quality and almost no clickbait
bzmrgonz · 1h ago
not to mention that soothing roller coaster voice of his!! hahaha.
umanwizard · 1h ago
> and why you must too
No, I don't think I do. Been working great for me so far.
scellus · 1h ago
If one wants to follow AI development mostly in the sense of LLMs and associated frontier models, that's an excellent list with over half of the names familiar, to whom I have converged independently.
I have a list in X for AI; it's the best source of information overall on the subject, although some podcasts or RSS feeds directly from the long-form writers would be quite close. (If one is a researcher themselves, then of course it's a must to follow the paper feeds, not commentary or secondary references.)
I'd add https://epoch.ai to the list, on podcasts at least Dwarkesh Patel; on blogs Peter Wildeford (a superforecaster), @omarsar0 aka elvis from DAIR in X, also many researchers directly although some of them like roon or @tszzl are more entertaining than informative.
The point about polluted information environment resonates on me; in general but especially with AI. You get a very incomplete and strange understanding by following something like NYT who seem to concentrate more on politics than technology itself.
Of course there are adjacent areas of ML or AI where the sources would be completely different, say protein or genomics models, or weather models, or research on diffusion, image generation etc. The field is nowadays so large and active that it's hard to grasp everything that is happening on the surface level.
Do you _have_ to follow? Of course not, people over here are just typically curious and willing to follow groundbreaking technological advancements. In some cases like in software development I'd also say just skipping AI is destructive to the career in the long term, although there one can take a tools approach instead of trying to keep track of every announcement. (My work is such that I'm expected to keep track of the whole thing on a general level.)
reactordev · 2h ago
Andrej's talks have helped me tremendously. They're up on YT [0]. For a long time I used to mentor and help machine learning scientists but when I hear Andrej speak, it's like I'm the student without any knowledge. It was a really strange feeling at first but I've come to value so much. I'm Jon Snow. I know Nothing (compared to Andrej).
The big question for me at the moment is whether to go pay-for-tool or pay-for-model.
If I pay for a tool that includes access to frontier models, then they'll keep the models up to date over time for me, let me use models from multiple providers, and the tool is carefully designed around the capabilities and limitations of the models it works with. On the other hand I can't really use the powerful model the tool works with for other applications or write my own.
If I pay for models, then I can only really use it with that manufacturers tools or tools that aren't optimised for the model but allow you to bring your own keys, and if the model provider I'm paying falls behind then I'm tied in for the duration of the contract. The big advantage is that there is a lot of innovation in tooling happening at the moment and you can avoid being locked out of that or having to pay many times for access to the same frontier models accross multiple different tools.
lumost · 2h ago
I think we'll see a shift to BYO LLM in the future. I saw that overleaf offers an assistant feature and an associated subscription. But I already pay for ChatGPT and Github copilot, I'm not going to pay for a third assistant - particularly when I can just bypass overleaf and use copilot in vscode directly.
smarx007 · 1h ago
I think Github already offers BYOK (bring your own key). I think the problem is that you are still expected to pay for all the premium subs, the key only allows you to go above the rate/monthly limits.
I think Jetbrains does it better with a full BYOM for models, including ollama. And I think if you go for ollama, you only need to pay for the IDE license, not for the AI add-on but don't quote me on that.
kybernetikos · 2h ago
I think commercially there'll be a really strong pull for dev tools to resell access to frontier models.
It's hard to make money out of dev tools, but if you tie in a service people are prepared to pay hundreds of dollars a month for, then suddenly it looks easier to make money out of an otherwise unsaleable IDE plugin that accesses that service.
vunderba · 1h ago
Yeah, Raycast and Warp are two tools that are definitely trying to do this.
vunderba · 1h ago
For me the issue with paying for a tool (for example cursor) is that our incentives are not necessarily aligned.
Since I'm not bringing my own API key, it's in their best interest to either throttle my usage by slowing it down or subtly downgrading me to a smaller LLM behind the scenes.
n8cpdx · 2h ago
Article does not make the case for why you must keep up.
AFAICT the best strategy would have been to completely tune out AI for the last ~3 years:
- AI has not meaningfully improved productivity (unless you’re doing something super basic like react and were already bad at it). If you are using AI in a transformative way, that looks different today than it did 6 months ago.
- AI has not stolen jobs (end of ZIRP did that)
- The field changes so fast that you could completely tune out, and at any moment become up-to-date because the news from 3 months ago is irrelevant.
I don’t get where this meme that “you have to keep up” comes from.
You have agency. You can get off the treadmill. You will be fine.
bluefirebrand · 1h ago
I have been tuning out for the last ~3 years and unfortunately it hasn't been the best strategy because the hype is still running roughshod all over me
It is very likely my employer will use my AI apathy as an excuse to include me in the next round of layoffs, compared to my coworkers that are very AI enthusiastic
chasd00 · 37m ago
They're not that hard to learn how to use. If your employer asks you to use one then just learn whatever tool the license you for, it's not that big of a deal. It's like learning how to use your employer's official IDE or email client.
scellus · 1h ago
Maybe that would even be justified, if you are a software developer? If not now, at least soon.
Imagine a software developer who refuses to use IDEs, or any kind of editor beyond sed, or version control, or some other essential tool. AI is soon similar, except in rare niche cases.
hagbarth · 1h ago
Version control is different since it’s collaboration with the rest of the org.
The rest: if they are just as productive as others, I would not care one bit. Tool use as a metric is just bad.
sshine · 1h ago
I've met many people smarter than myself who object to all of those.
One professor thought syntax highlighting was a distraction.
Lots of colleagues used vim/helix instead of IDEs.
I haven't met anyone who refused version control from an intelligent standpoint.
The most reasonable objection to AI from people who don't hate it are:
I just don't know how it could help me; it's not as skilled as me at my job, and I'm already doing fine.
sshine · 1h ago
The article uses "why" in the title, but does not follow through with an answer.
It sort of hints at one reason:
> The most common errors of misunderstanding are either underestimation (“it’s all hype that will blow over”) or overestimation (“I don’t need programmers anymore”). These patterns are rooted in a lack of a solid understanding of the technology and how it is evolving over time.
So if you don't at least find some middle ground between those two poles, you will make uninformed choices.
But I agree: It is safe to ignore AI for now.
I do sense that some people attach to AI because of a fundamental anxiety that it might transform society quickly and detrimentally, because that's part of the hype speech ("it will murder us all, it will make us all unemployed, it will turn us into slaves, maybe you can ride the dragon, and maybe you must").
---
> AI has not meaningfully improved productivity
This is contended.
As the article says, we are in one of the most polluted information environments.
People will say "It's absolutely useless" and "It has fundamentally changed my life."
So neither extreme can be taken at face value as representative; they're samples of a murky picture.
> The field changes so fast that you could completely tune out
It's not that fast, in my opinion. Last big steps:
- Transformer architecture (2017)
- Larger models with greater performance (2020-)
- Chain of thought (research in 2022, commercial breakthrough in 2024)
- Agents (since forever, but 2022 for GPT-based agentic frameworks)
Other things happened; for example, DeepSeek making an architectural breakthrough and challenging the financial model of open/closed weights.
But most of the hype is just people trying to make commercial success on a few cornerstone breakthroughs.
In one to two years, maybe we can add one more major advancement.
Don’t you ever let people make you think you have to tirelessly follow some tech progress or stay on top of things all the time.
Progress is like a bus. You can just get on board at any time. You’re not going to fall behind. And staying up to date doesn't keep you “ahead” of anyone.
Doing things is what gets you ahead, and if you don’t feel like doing something right now, don’t worry about, do something later, and you’ll be ahead of people who aren’t doing anything at that moment.
nice_byte · 1h ago
Doesn't explain the "why you must too" part
linsomniac · 1h ago
Here's my current rule of thumb: If you have successfully built a couple projects using agentic tooling and Claude 4 or similar models: you are doing a fine job of keeping up. Otherwise, you are at least a generation behind.
bluefirebrand · 1h ago
Behind what?
Isn't the whole promise of AI tools that they just work?
What skill am I missing out on learning, exactly, by not using them right now? Prompt Engineering?
I think I'm a reasonably good communicator in both voice and text, so what skill am I failing to train by not using LLMs right now?
linsomniac · 1h ago
>Isn't the whole promise of AI tools that they just work?
No, not at all. Like with pretty much any development tool, you need to get proficient with them.
>what skill am I missing out on
At this point, it seems like pretty much all of them related to generative AI. But, the most recent of them that I'll point at is: tooling tooling tooling, and prompting. But the specific answer (to answer your "exactly") is going to depend on you and what problems you are solving. That's why on tries not to fall behind, so you can see how to use tooling in a rapidly evolving landscape, for your exact circumstances.
>I think I'm a reasonably good communicator in both voice and text, so what skill am I failing to train by not using LLMs right now?
You know how to achieve something you will use different words with different people? You don't talk to your spouse the same way you talk to your parents or your children or your friends or your coworkers, right? You understand that if you are familiar with someone you speak to them differently if you want to achieve something, yes?
dingnuts · 1h ago
this is just ridiculous. you can get up to speed with SOTA tooling in a few hours. A system prompt is just a prompt that runs every time. Tool calls are just patterns that are fine tuned into place so that we can parse specific types of LLM output with traditional software. Agents are just a LLM REPL with a context specific system prompt, and limited ability to execute commands
none of this stuff is complicated, and the models themselves have been basically the same since GPT-2 was released years ago
chasd00 · 22m ago
> A system prompt is just a prompt that runs every time. Tool calls are just patterns that are fine tuned into place so that we can parse specific types of LLM output with traditional software. Agents are just a LLM REPL with a context specific system prompt, and limited ability to execute commands
pulling the covers back so hard and so fast is going to be shocking for some.
To make it more concrete you can try and build something yourself. Grab a small model off of hugging face that you can run locally. Then put a rest API in front of it so you can make a request with curl, send in some text, and get back in the response what the llm returned. Now in the API prepend some text to what came on the request ( this is your system prompt ) like "you are an expert programmer, be brief and concise when answering the following". Now add a session to your API and include the past 5 requests from the same user along with the new one when passing to the llm. Update your prepended text (the system prompt) with "consider the first 5 requests/responses when formulating your response to the question". you can see where this is going, all of the tools and agents are some combination of the above and/or even adding more than one model.
At the end of the day, everyone has a LLM at the core predicting and outputting the next most likely string of characters that would follow from an input string of characters.
Disposal8433 · 1h ago
Your FOMO threats are way too obvious, and I'm not falling for it . I'm behind you in C++ because I don't use AI? That's ridiculous.
linsomniac · 1h ago
It may not be as ridiculous as you think. I have 25 years of experience with Python, and the generative AI tooling is teaching me useful things in Python when I work with it.
Despite the feeling that it's a fast-moving field, most of the differences in actual models over the last years are in degree and not kind, and the majority of ongoing work is in tooling and integrations, which you can probably keep up with as it seems useful for your work. Remembering that it's a model of text and is ungrounded goes a long way to discerning what kinds of work it's useful for (where verification of output is either straightforward or unnecessary), and what kinds of work it's not useful for.
I also have not experienced the post's claim that: "Generative AI has been the fastest moving technology I have seen in my lifetime." I can't speak for the author, but I've been in this field from when "SVMs are the new hotness and neural networks are a joke!" to the entire explosion of deep learning, and insane number of DL frameworks around the 20-teens, all within a decade (remember implementing restricted Boltzmann machines and pre-training?). Similarly I saw "don't use JS for anything other than enhancing the UX" to single page webapps being the standard in the same timeframe.
Unless someone's aim is to be on that list of "High signal" people, it's far better to just keep your head down until you actually need these solutions. As an example, I left webdev work around the time of backbone.js, one of the first attempts at front end MVC for single pages apps. Then the great React/Angular wars began, and I just ignored it. A decade later I was working with a webdev team and learned React in a few days, very glad I did not stress about "keeping up" during the period of non-stop changing. Another example is just 5 years ago everyone was trying to learn how to implement LSTMs from scratch... only to have that model essentially become obsolete with the rise of transformers.
Multiple times over my career I've learned lesson that moving fast is another way of saying immature. One would find more success learning about the GLM (or god forbid understanding to identify survival analysis problems) and all of it's still under appreciated uses for day-to-day problem solving (old does not imply obsolete) than learning the "prompt hack of the week".
- 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.
A lot of people are feeling HN is saturated with AI posts whether it is how MCP is like USB-C (repeated so much you know it is NPCs) or how outraged people are that their sh1t fanfics are being hoovered up to train AI.
This piece is not “news”, it’s a summary which is tepid at best, I wish people had some better judgement about what they vote up.
The minutiae of how next token prediction works is rarely appreciated by lay people. They don't care about dot products, or embeddings, or any of it. There's basically no advantage to explaining how that part works since most people won't understand, retain, or appreciate it.
Last week I showed some colleagues how to do some basic things with Claude Code and they were like "wow, I didn't even know this existed". Bro, what are you even doing.
There is definitely a lot of hype and the lunatics on Linkedin are having a blast, but to put it mildly I don't think it's a bad investment to experiment a bit with what's possible with the SOTA.
The trouble is that the advice in the post will have very little impact on "understanding how LLMs work". The number of people who talk about LLMs daily but have never run an LLM local, and certainly never "opened it up to mess around" is very large.
A fun weekend exercise that anyone can do is to implement speculative decoding[0] using local LLMs. You'll learn a lot more about how LLMs work than reading every blog/twitter stream mentioned there.
0. https://research.google/blog/looking-back-at-speculative-dec...
It is ridiculously cool, but I think anybody developer who is out of the loop could easily get back into the loop at any moment without having to stay caught up most of the time.
The problem I see is that a lot of people are grossly misaligned with the state of the art, and it does take a bit of experimentation to understand how to work with an LLM. Even basic stuff like how to work with context isn't immediately obvious.
There's far too much noise, churn, and indecision at this stage to get any productive value out of riding the bleeding edge.
If it's actually revolutionary, you'll hear about it on HN.
[1] https://x.com/burkov
It's fine to go do other things with your precious time instead.
Unless you are building an AI startup furiously searching for PMF before your runway expires, I don't see the urgency.
What is an "AI startup"? If you add a chatbot do your product, are you an "AI startup"? Does "startup" require having no moat? Can you be an established stable business that loses everything to a startup that leapfrogs your tech, AltaVista?
https://newsletter.pragmaticengineer.com/ https://substack.com/@garymarcus
Plus I subscribe to updates from Python packages like Langchain and PydanticAI to see what they're up to, since that's usually reflective of what's happening in broader industry.
I'm not on X anymore so I can't follow many of the folks directly, but some of them (like Simon Willison) also post on BlueSky and Mastodon, fortunately. Some folks like Sebastian Raschka and Chip Huyen also post on LinkedIn. Kind of all over, but eventually, I see a good amount of what's happening.
I'm currently not using AI or LLMs in any of my day-to-day work.
Yeah it's not a fad, but I think it's really not as useful to me right now as the hype seems to suggest
I'm going to keep an eye on developments, but I'm not using it for my day to day either. I'm just not seeing the same value other people seem to be seeing right now and I'm not going to exhaust myself trying to keep up
One day Claude is the best. Next is Cursor. People are switching tools every two weeks trying to keep up
Not for me thanks. Especially not with how inconsistent the quality and output still are
btw here's a link to the karpathy videos https://karpathy.ai/zero-to-hero.html
edit: i use claude and chatgpt day to day to help with simple things like regex, a replacement for google search in same cases, and self contained classes, functions, and other small discreet blocks of code.
I've also got some gems from Microsofts Build talks, specifically whenever Scott Hanselman and Mark Russinovich get together, e.g.: https://www.youtube.com/watch?v=KIFDVOXMNDc
Happily canceled my GPT Plus this week; personally not gonna feed that beast any longer! As well it can not generate maps (create road trip travel maps showing distance between locations to share with friends, a creek tubing map route & etc) at all like Gemini can for free.
This would increase employment ceteris paribus. That's like saying inventing new programming languages is bad because they're too productive.
https://en.wikipedia.org/wiki/Jevons_paradox
> This would increase employment ceteris paribus.
This might be true, but if it is, the one "doing the job ten use[d] to" would not actually being doing the same kind of work at all, and so therefore might not be the same people or even same kind of people. Even if we do Jevons ourselves out of this situation, it might still spell employment disaster for mid level coders, while increasing employment for product managers.
So You and all other people like to save money are going to continue spend the same thousands on such a resource when AI can do what they do in a few minutes or more for WAY LESS? UX Design was supposedly a growing field ... not at anymore! Definitely one can do the same thing in that field that 10 did.
Further, future mobile AI devices will pull the information and put it all on the lock screen of your AI device visualizing the data in a fun new way. Technology that makes things simpler and more magical get adopted yet visits to websites will significantly decline.
For federal workers who have lost their jobs they are feeling this pain competing for jobs against each other and now AI. It will only get worse for designers because it's now cheaper and faster to use AI to design logos, sites, apps to even including do vibe coding for the front end development to possibly the backend but that's not my specialty yet no doubt I vibe coded front-ends.
Is this not obvious?
Why do people hide behind this ridiculous analogy: "That's right, there are no carpenters or lumberjacks anymore because power tools were invented"
???
I mean sure the analogy is catchy and makes surface level sense, but can your brain do some analysis outside the context of an analogy??? It makes no sense that all of AI can be completely characterized by an analogy that isn't even accurate yet people delusionally just regurgitate the analogy most fitting with the fantasy reality they prefer.
Are you saying that when the sewing machine was invented, it would be in the employees interest to not learn how to use it? Or when the computer was invented, it’s not in the employees interest to learn how to use it?
Even if you are a software engineer and are fired / laid off / pushed out of the industry because of AI, knowing how to use AI, its risks, etc is still helpful. It’s a skill you can hopefully use in your next career, whatever you pivot to. Unless you pivot to manual labor.
Thinking otherwise is shortsighted.
No, I don't think I do. Been working great for me so far.
I have a list in X for AI; it's the best source of information overall on the subject, although some podcasts or RSS feeds directly from the long-form writers would be quite close. (If one is a researcher themselves, then of course it's a must to follow the paper feeds, not commentary or secondary references.)
I'd add https://epoch.ai to the list, on podcasts at least Dwarkesh Patel; on blogs Peter Wildeford (a superforecaster), @omarsar0 aka elvis from DAIR in X, also many researchers directly although some of them like roon or @tszzl are more entertaining than informative.
The point about polluted information environment resonates on me; in general but especially with AI. You get a very incomplete and strange understanding by following something like NYT who seem to concentrate more on politics than technology itself.
Of course there are adjacent areas of ML or AI where the sources would be completely different, say protein or genomics models, or weather models, or research on diffusion, image generation etc. The field is nowadays so large and active that it's hard to grasp everything that is happening on the surface level.
Do you _have_ to follow? Of course not, people over here are just typically curious and willing to follow groundbreaking technological advancements. In some cases like in software development I'd also say just skipping AI is destructive to the career in the long term, although there one can take a tools approach instead of trying to keep track of every announcement. (My work is such that I'm expected to keep track of the whole thing on a general level.)
[0] https://www.youtube.com/@AndrejKarpathy
If I pay for a tool that includes access to frontier models, then they'll keep the models up to date over time for me, let me use models from multiple providers, and the tool is carefully designed around the capabilities and limitations of the models it works with. On the other hand I can't really use the powerful model the tool works with for other applications or write my own.
If I pay for models, then I can only really use it with that manufacturers tools or tools that aren't optimised for the model but allow you to bring your own keys, and if the model provider I'm paying falls behind then I'm tied in for the duration of the contract. The big advantage is that there is a lot of innovation in tooling happening at the moment and you can avoid being locked out of that or having to pay many times for access to the same frontier models accross multiple different tools.
I think Jetbrains does it better with a full BYOM for models, including ollama. And I think if you go for ollama, you only need to pay for the IDE license, not for the AI add-on but don't quote me on that.
It's hard to make money out of dev tools, but if you tie in a service people are prepared to pay hundreds of dollars a month for, then suddenly it looks easier to make money out of an otherwise unsaleable IDE plugin that accesses that service.
Since I'm not bringing my own API key, it's in their best interest to either throttle my usage by slowing it down or subtly downgrading me to a smaller LLM behind the scenes.
AFAICT the best strategy would have been to completely tune out AI for the last ~3 years:
- AI has not meaningfully improved productivity (unless you’re doing something super basic like react and were already bad at it). If you are using AI in a transformative way, that looks different today than it did 6 months ago. - AI has not stolen jobs (end of ZIRP did that) - The field changes so fast that you could completely tune out, and at any moment become up-to-date because the news from 3 months ago is irrelevant.
I don’t get where this meme that “you have to keep up” comes from.
You have agency. You can get off the treadmill. You will be fine.
It is very likely my employer will use my AI apathy as an excuse to include me in the next round of layoffs, compared to my coworkers that are very AI enthusiastic
Imagine a software developer who refuses to use IDEs, or any kind of editor beyond sed, or version control, or some other essential tool. AI is soon similar, except in rare niche cases.
The rest: if they are just as productive as others, I would not care one bit. Tool use as a metric is just bad.
One professor thought syntax highlighting was a distraction.
Lots of colleagues used vim/helix instead of IDEs.
I haven't met anyone who refused version control from an intelligent standpoint.
The most reasonable objection to AI from people who don't hate it are:
It sort of hints at one reason:
> The most common errors of misunderstanding are either underestimation (“it’s all hype that will blow over”) or overestimation (“I don’t need programmers anymore”). These patterns are rooted in a lack of a solid understanding of the technology and how it is evolving over time.
So if you don't at least find some middle ground between those two poles, you will make uninformed choices.
But I agree: It is safe to ignore AI for now.
I do sense that some people attach to AI because of a fundamental anxiety that it might transform society quickly and detrimentally, because that's part of the hype speech ("it will murder us all, it will make us all unemployed, it will turn us into slaves, maybe you can ride the dragon, and maybe you must").
---
> AI has not meaningfully improved productivity
This is contended.
As the article says, we are in one of the most polluted information environments.
People will say "It's absolutely useless" and "It has fundamentally changed my life."
So neither extreme can be taken at face value as representative; they're samples of a murky picture.
> The field changes so fast that you could completely tune out
It's not that fast, in my opinion. Last big steps:
Other things happened; for example, DeepSeek making an architectural breakthrough and challenging the financial model of open/closed weights.But most of the hype is just people trying to make commercial success on a few cornerstone breakthroughs.
In one to two years, maybe we can add one more major advancement.
https://x.com/rowancheung - Rowan Cheung: “Daily” updates and insider access
Progress is like a bus. You can just get on board at any time. You’re not going to fall behind. And staying up to date doesn't keep you “ahead” of anyone.
Doing things is what gets you ahead, and if you don’t feel like doing something right now, don’t worry about, do something later, and you’ll be ahead of people who aren’t doing anything at that moment.
Isn't the whole promise of AI tools that they just work?
What skill am I missing out on learning, exactly, by not using them right now? Prompt Engineering?
I think I'm a reasonably good communicator in both voice and text, so what skill am I failing to train by not using LLMs right now?
No, not at all. Like with pretty much any development tool, you need to get proficient with them.
>what skill am I missing out on
At this point, it seems like pretty much all of them related to generative AI. But, the most recent of them that I'll point at is: tooling tooling tooling, and prompting. But the specific answer (to answer your "exactly") is going to depend on you and what problems you are solving. That's why on tries not to fall behind, so you can see how to use tooling in a rapidly evolving landscape, for your exact circumstances.
>I think I'm a reasonably good communicator in both voice and text, so what skill am I failing to train by not using LLMs right now?
You know how to achieve something you will use different words with different people? You don't talk to your spouse the same way you talk to your parents or your children or your friends or your coworkers, right? You understand that if you are familiar with someone you speak to them differently if you want to achieve something, yes?
none of this stuff is complicated, and the models themselves have been basically the same since GPT-2 was released years ago
pulling the covers back so hard and so fast is going to be shocking for some.
To make it more concrete you can try and build something yourself. Grab a small model off of hugging face that you can run locally. Then put a rest API in front of it so you can make a request with curl, send in some text, and get back in the response what the llm returned. Now in the API prepend some text to what came on the request ( this is your system prompt ) like "you are an expert programmer, be brief and concise when answering the following". Now add a session to your API and include the past 5 requests from the same user along with the new one when passing to the llm. Update your prepended text (the system prompt) with "consider the first 5 requests/responses when formulating your response to the question". you can see where this is going, all of the tools and agents are some combination of the above and/or even adding more than one model.
At the end of the day, everyone has a LLM at the core predicting and outputting the next most likely string of characters that would follow from an input string of characters.