I can feel the cold wind of the next AI winter coming on. It's inevitable. Computers are good at emulating intelligent behavior, people get excited that it's around the corner, and the hype boils over. This isn't the last time this will happen.
On all corners people work on so many small pieces advancing what we have.
And plenty of obvious things are not here yet like a full local dev cycle aka ai uses the IDE to change code them executes it, fixes compiler issues and unit tests and then opens a pr.
Local agents/ agents having secure and selected access to our data like giving my agent read only access to my bank account and a 2factor way for it to send money.
Deepseek's reinforcement learning is also a huge new lead.
Nonetheless in parallel robots are coming too.
GenAI is getting better and better. Faster and better video and cheaper. 3D meshes, textures first GenAI ads
In reality people doing "menial" jobs are smart and learn and operate with a lot of nuance than people ignore given unfamiliarity or just prejudice. Do you prefer to talk to a chatbot or a real human when you have a problem, how confident are you really, that even if the bot knows what the problem is it would be able to solve it.
Lots of problems with customer care is anchored in the issue that support staff is not allowed to fix or resolve problems without escalation or attempts at keeping you from costing more money. The bot might be better at it for the company because it will frustrate you enough to give up that 30 bucks refund, idk.
Ai seems to change a lot the dynamics of corporate jobs but I haven't seen yet anything that would be a game changer outside of it. Its great for searching company unorganised and messy knowledge bases.
>Computers are good at emulating intelligent behavior
Which implies that the brain is some kind of transcendent device that can backdoor physics to output incredible intelligence unique to it's magical structure.
Maybe LLMs aren't the key, but as far as we can tell the brain is also just another computer.
And let me be very clear on why, because I love having conversations about this theme: it promises to be an adversarial and frustrating exchange.
I have found similar when giving backstory and needing help to start structuring difficult conversations where I want to say the right thing but also need to be sensitive.
> she watched a operation that failed for me for entertainment, for instance
You make your own choices, but cutting a person like this off would be very reasonable
For coding, I use cursor composer to gather context about the existing codebase (context.md). Then I paste that into DeepSeek R1 to iterate on requirements and draft a high level design document, maybe some implementation details (design.md).
Paste that back into composer, and iterate; then write tests. When I'm almost done, I ask composer to generate me a document on the changes it made and I double check that with R1 again for a final pass (changes.md).
Then I'm basically done.
This is architect-editor mode: https://aider.chat/2024/09/26/architect.html.
I've found Cursor + DeepSeek R1 extremely useful, to the point that I've structured a lot of documents in the codebase to be easily greppable and executable by composer. Benefit of that is that other developers (and their composers) can read the docs themselves.
Engineers can self-onboard onto the codebase, and non-technical people can unstuck themselves with SQL statements with composer now.
Seriously, this is the only useful flow I've found for AI coding in general..
For design.md, I have a prompt because we're now working with a reasoning model and doing structured data extraction: create me an issue on Linear with a title, description, and a to-do list.
I would recommend trying the approach yourself and saving the prompts if you can nail down the repetitive asks.
Chat function is just chat; it can't edit your files.
Composer probably relies on prompt engineering to do editor-architecture, as it reads and writes to your codebase. But it's heavily heavily to Sonnet 3.5 and tool-calling.
For architecture-type stuff, I prefer DeepSeek R1 as reasoning models do better on high level design. Which is why I will copy and paste in and out of compsoer.
- Discussing the various stages of candymaking and their relation to the fundamental properties of sugar syrups, and which candies are crystalline vs amorphous. It turns out junior mints are fudge. Fondant is really just fudge. Everything is fudge, my god.
- Summarizing various SEC filings and related paperwork to understand the implications of an activist investor intervening in a company
- Discussing the relative film merits of the movie Labyrinth and other similar 80s kitsch movies. ChatGPT mentioned the phenomenon of "twin films" which was an interesting digression.
- Learning about various languages Tolkien invented and their ties to actual linguistics of natural languages and other conlangs
- Some dimensional analysis of volumes, specifically relating to things like "how many beans are in the jar" estimation and what the min and max of a particular weight of coins might be valued, in terms of both a par value based on a standard coin mix and outliers of, for example, old dimes that are pure silver.
- Discussion of quines in prolog and other languages, which resulted in a very interesting ChatGPT bug where it started recursing and broke when trying to write a prolog quine.
- Back of the envelope economic calculations around the magnitude of the housing deficit and the relative GDP cost for providing enough housing quickly enough to make an impact. Spoiler: it's probably unreasonably expensive to build enough houses to bring down housing prices by any significant degree, and even if we wanted to, there's not enough skilled workers.
- A number of podcasts transcribed. (I hate audio and meandering, so transcribed and summarized is perfect) I could use whisper and a python script to do this, but I'd rather let ChatGPT do the legwork, and it actually used a more modern model and method of processing than I would have naively used.
I find Github Copilot to be a really great autocomplete. I frequently write the comment at the top of a function and hit tab and it writes the whole function. This is dependent on typescript and having a relatively standard codebase but I think those things are useful on their own. You really have to limit it in terms of scope and specifics, but it lets me think high level instead of worrying about syntax.
Fudge is made with milk - am I missing a joke?
For example asking it something like "I have an elixir app that is started with `mix ...` can you give me a Dockerfile to run it in a container?"
It can also do things like "Given this code snippet, can you make it more Pythonic" or even generate simple apps from scratch.
For example, a prompt like "Can you write me a snake game in HTML and JavaScript? The snake should eat hot dog emojis to grow longer." will actually generate something that works. You can see the generated code for that prompt at https://claude.site/artifacts/34540f88-965e-45ca-8083-040e30...
Following up with "Can you make it so that people can swipe on mobile to control the snake?" generates https://claude.site/artifacts/651e957a-9957-488c-ae6b-e81348... which is pretty good IMO for 30 seconds of effort.
It also has a surprisingly competent analysis mode where you can upload a CSV and have it generate charts and analyze the data.
It's not perfect, it'll sometimes get confused or generate some dubious code, but you can quickly get to a 90% good solution with 1% of the effort, which is pretty impressive IMO.
One thing I can't figure out how to get LLMs to do is truely finish work. For example if I have 100 items that need xyz done to them, it will do it for the first 10 or so and say ~"and so on". I have a lot of trouble getting LLMs to do tasks that might take 10 mins - 1h. They always seem to simply want to give an example. Batch processing is the answer, I guess, or perhaps more 'agentic' models/tools - but I wonder if there are other ways.
(Other answers are people gathering examples, which is nice, but I'm looking for more structured things.
And, I suppose I could ask an LLM, but my main problem is that... I don't really _trust_ LLMs yet :D )
It (mostly) exceeds and excels at every task I use it for. I'm rarely disappointed. YMMV.
Absolutely life-changing for me.
It also helps me getting started with new content, kind of building the scaffolding of, say, a blog or social post. It still needs adaption and fine-tuning, but getting rid of a white page is a great help for me.
And I use LLMs to play through ideas and headlines. I would normally do this with other humans, but since working full remote, its a nice sparing partner, although the AI not being able to really give criticism is a bit annoying.
The tools also make it easier to write in English as a non-native, making sure my text does not include any false friends or grammar errors.
I’ve managed to get ChatGPT to a good place in this regard recently and it’s better for it. In fact, it’s been doing such a good job recently that it almost seems like…human like.
I’ll have to look at my prompts, but somehow I got it from licking my ass and telling me everything I say and do is amazing to a point now where it almost seems eager to tell me I’m wrong.
Which is fantastic, huge improvement. I don’t really use it for coding though, because I am not a programmer. I would have no means today to correctly evaluate 90% of what it would return me.
LLMs are pretty good at translation between human languages which makes sense since they are language models after all. They are better at this any any other technology.
The state of the art image ones can also probably do OCR and handwriting recognition better than any other software though may be expensive to run in large volume. But if you need to take picture of a notebook page with your camera phone an LLM can quickly OCR it.
I myself use them a lot, though I constantly feel that I would be able to get more out of them if only I were smarter.
we import descriptions of products from a seller. the problem is they are mental ( probably written by chatgpt :)) and are way too long. we need only small blurb.
I put our style guide and given text to chatgpt and I get somehow reasonable description back. then editors still need to check it, but it's way less work.
* in panel A, some guy is proud to use ChatGPT to turn 3 lines of text into a 10 pages report
* in panel B, some girl is happy to use ChatGPT to summarize the 10 pages report into 3 lines
It was meant to be _satire_, not the sales pitch ;)
> ZIZEK: that AI will be the death of learning & so on; to this, I say NO! My student brings me their essay, which has been written by AI, & I plug it into my grading AI, & we are free! While the 'learning' happens, our superego satisfied, we are free now to learn whatever we want
I find it good for complex SQL, reviewing emails, and Godot assistance (I'm a beginner game Dev).
There are also times when I have programming questions and I might try to use chatgpt, with mixed results.
Our company has tried to integrate it into one of our products, and I find it troubling how on occassion it is confidently giving bad results, but my concern seems to be in the minority.
EDIT: there was also a large refactor I did recently which involved lots of repeatable, but not super regexable, changes - chatgpt forgot where it was as I went through it, but other than working around that it was very useful.
I'll give two recent use-cases that may provide a hint of their ultimate utility:
1) I've been modernising 2010-era ASP.NET code written by former VB programmers that looooved to sprinkle try { ... } catch( Exception e ) { throw e; } throughout. I mean thousands upon thousands of instances of these pointless magical incantations that do nothing except screw up stack traces. They probably thought it was the equivalent of "ON ERROR RESUME NEXT", but... no, not really. Anyway, I asked ChatGPT in "Reasoning" mode to write a CLI tool utilising the Roslyn C# compiler SDK to help clean this up. It took about three prompts and less than an hour, and it spat out 300 lines of code that required less than 10 to be modified by me. It deleted something like 10K lines of garbage code from a code base for me. Because I used a proper compiler toolkit, there was no risk of hallucinations, so the change Just Worked.
2) I was recently troubleshooting some thread pool issues. I suspect that some long-running requests were overlapping in time, but Azure's KQL doesn't directly provide a timeline graphical view. I dumped out the data into JSON, gave ChatGPT a snippet, and told it to make me a visualiser using HTML and JS. I then simply pasted in the full JSON dump (~1 MB) and ta-da instant timeline overlap visualiser! It even supported scrolling and zooming. Neat.
1. Exploring a new domain and getting some terms I can google for.
2. Making small scripts to do things like query github's GraphQL API.
3. Autocomplete of code using copilot.
It's a weak jack of all trades: it knows a fair amount about the sum of human knowledge (which is objectively super-human), but can't go deep on any one thing, and still seriously lags behind humans in terms of reasoning. It's an assistant that all book smarts and no street smarts. Or maybe: it's a search engine for insanely specific things.
Rote work, as well. Things like porting an enum from one programming language to another: past the source language into a comment and start it off with one or two lines in the target language. Dozens of tabs are surely faster than manual typing, copy paste, or figuring out vim movements/macros.
The answer staring the OP right in the face.
> You've allowed the AI to train you to use the wrong word through a shared delusion, haven't you.
What an awful interpretation, phrased in the most demeaning manner possible. No, I double check everything the AI suggests. That's basic competency with the things.
> Me: I'm trying to use a cooperator in my schematic, how do I hook it up? Also what's a suitable cooperator for 3.3v logic level?
> Gemini: It sounds like you might be thinking of a comparator, not a "cooperator," in your schematic. Comparators are electronic circuits that compare two voltages [...] LM393: This is a low-power, dual comparator that can operate from a 3.3V supply. Its output can be directly connected to most 3.3V logic circuits.
They might be able to help stitch together modules (like sensor boards plugged into microcontrollers) and definitely can write code to get things going, but they fall flat on their face hard for any kind of bare bones electronics design. Like 5% success rate and 95% totally incorrect hallucinations.
The models can also output code that can be turned into a schematic through an interpreter, but there is virtually zero training data for this because humans always use and work with schematics.
Most recently I tried to use them both to solve a programming problem that isn't well documented in the usual channels (Reddit, StackOverflow, etc) and found it to be quite a disappointing and frustrating experience. It just constantly, enthusiastically fed me total bullshit, with functions that don't exist or don't do what the LLM seems to "think" they do. I'm sure I'm just "holding it wrong" but my impression at this stage is that it is only capable of solving problems that are trivially solvable using a traditional search engine, with the added friction that if the problem isn't trivially solvable, it won't actually tell you that but will waste your time with non-obvious wrong answers.
I did have a slightly more positive experience when asking it about various chess engine optimisation algorithms. I wasn't trying to use the code it generated, just to better understand what the popular algorithms are and how they work. So I think they might work best when there is an abundance of helpful information out there and you just don't want to read through it all. Even then, I obviously don't know what ChatGPT was leaving out in the summary it provided.
For example, you have a plant you can't identify. Hard to Google search with words. "Plant with small red berries and...". You could reverse image search your photo of it, probably won't help either. Show an LLM the photo (some accept images now). LLM tells you what it thinks. Now you Google search "Ribes rubrum" to verify it. Much easier.
You've got a complicated medical problem that's been going on for months. A google search of all the factors involved would be excessively long and throw up all sorts of random stuff. You describe the whole scenario to an LLM and it gives you four ideas. You can now search those specific conditions and see how well they actually match.
I've found there are actually a lot of questions that fit in that sort of NP complexity category.
Same, it's good for repetitive things, things that have been answered 1000 times on stack overflow, translations, but that's about it. If you work on anything remotely new/hard it's mostly disappointing, you have to babysit it every step of the way and rewrite most of what it's shitting out in the end anyways.
I think it just made it obvious that 90% of tech jobs basically amount to writing the same CRUD thing over and over again & mobile/web apps with very common designs and features.
Then I had a better idea: I spent 20 minutes baby wearing, walking and dictating everything about my startup to ChatGPT. Later I took all that text and labeled it as a brain dump, plus my product support portal and some screenshots of my marketing material. Gave it all to ChatGPT again and asked it to answer each of the questions in the form. That's it. I have a pretty good version 1 which I can revise today and be done with it.
Many, many hours saved. I have tens of examples like that.
The product documentation I provided it with was also created with the help of GPT, and that saved me even more time.
- I have these three ingredients; recommend Italian main courses.
- What other ingredients pair well with this?
- How can I "level up" this dish if I want to impress?
- Can I substitute X for Y?
- Generate a family-friendly meal with lots of veggies using leftover roast chicken.
this is good enough sell for me, and it's like sub 1-in-50 that it's "auto-complete on mushrooms" (again my experience, YMMV).
An awful lot of the time, my day to day work involves writing one piece of code and then copy-pasting it changing a few variable names. Even if I factor out the code into a method, I've still got to call that method with the different names. CoPilot takes care of that drudgery and saves me countless minutes per day. It therefore pays for itself.
I also use ChatGPT every time I need some BASH script written to automate a boring process. I could spend 20-30 minutes searching for all the commands and arguments I would need, another 10 minutes typing in the script, another 10-20 minutes debugging my inevitable mistakes. Or I make sure to describe my requirements exactly (5-10 minutes), spend 5 minutes reviewing the output, iterate if necessary (usually because I wasn't clear enough in the instructions).
3-5x speed up for free. Who's not going to take that win?
Stop using Google search and use an AI. No more irrelevant results, no more ads. No more slop to wade through.
BTW I find Claude is great at making graphs and diagrams. If you pay ($20) you can hook it up to a local code base.
Huh? More like "slop exclusively generated for you", right? I have seen so much garbage answers from chat AIs.
* Figuring out where to start when learning new things (see also <https://news.ycombinator.com/item?id=43087685>)
One way I treat LLMs is as a "semantic search engine". I find that LLMs get
too many things wrong when I'm being specific, but they're pretty good at
pointing me in a general direction.
For example, I started learning about OS development and wanted to use Rust. I
used ChatGPT to generate a basic Rust UEFI project with some simple
bootloading code. It was broken, but it now gave me a foothold and I was able
to use other resources (e.g. OSDev wiki) to learn how to fix the broken bits.
It feels like a lot of software documentation isn't actually written for real
readers; instead being a somewhat arbitrary listing of a program's features.
When programs have this style of documentation, the worst case for figuring
out how to do a simple thing is reading the entire manual. (There are better
ways to write documentation, see e.g. <https://diataxis.fr/>)
One example is [gnuplot](http://www.gnuplot.info/). I wanted to learn how to
plot from the command line. I could have pieced together how to do it by
zipping around the
[gnuplot manual](http://www.gnuplot.info/docs_5.4/Gnuplot_5_4.pdf) and building
something up piecewise, but it was faster to instruct Claude directly. Once
Claude showed me how to do a particular thing (e.g. draw a scatter plot with
dots intstead of crosses) I then used the manual to find other similar
options.
Similar to the previous point. If I ask Claude to write a simple program using
a complex publicly-available API, it will probably write a broken program, but
it won't be *completely* bogus because it will be in the right "genre". It
will probably use some real modules, datatypes and functions in a realistic
way. These are often good leads for which code/documentation I should read.
I used this approach to write some programs that use the
[GHC API](https://hackage.haskell.org/package/ghc). There are hundreds of
modules, and when I asked Claude how to do something with the GHC API it wrote
relevant (if incorrect) code, which helped me teach myself.
My partner is Chinese and sometimes we talk about Chinese poetry. I'm not very
fluent in Chinese so it's hard for me the grasp the beauty in these poems.
Unfortunately literal English translations aren't very good. We've had some
success with asking LLMs to translate Chinese poems in the style of various
famous English poets. The translation is generally semantically correct, while
having a more pleasing use of the English language than a direct translation.As they state in the report, I don't think they can measure how many people just ignored the suggestion from claude right away ? Or delete more than half of the suggested code ?
Imagine if the real impact of AI is "suggesting things that people are discarding immediately ?"
Call the "Unamusing misuse of resource"... [1]
- Writing Python scripts to make charts out of Excel sheets, and then refine them. I could do it myself, but I would need to learn a library like Seaborn or similar which honestly is not especially intellectually stimulating, and then spend nontrivial amounts of time iterating on the actual code. With LLMs it's a breeze.
- Working with cumbersome LaTeX formatting, e.g. transposing a table, removing a column from a table, etc.
- Getting the tone just right in a professional email written in English to someone I don't know much (I'm not a native speaker so this is not trivial).
- Finding resources on topics that are tangential to what I do. For example, yesterday I needed to come up with some statistics on English words for a presentation I'm preparing, and I needed a free corpus where I could search for an n-gram and get frequencies of next words. I don't usually work with that kind of resource, it was just a one-off need. I asked for corpora of that kind and got a useful answer instantly. The manual process would probably have implied going through several options only to find that I needed a license or that they didn't provide the specific statistics I needed.
- Brainstorming on titles for scientific papers, presentations, names of concepts that you introduce on a paper, variable names, etc.
- Shortening a sentence in a paper that makes me go over the page limit, or polishing the English in a paragraph.
- Summarizing a text if I'm kind of interested in knowing the gist but have no time to read it whole.
- Answering quick questions on basic things that I forget, e.g. the parameters to make a Linux folder into a tar.gz. Man is too verbose and it takes time to sort the wheat from the chaff, Google is full of SEOd garbage these days and sometimes you need to skim a lot to find the actual answer, LLMs are much faster.
- Writing bureaucratic boilerplate, the typical texts with no real value but that you have to write (e.g. gender perspective statement on a grant request).
- Coming up with exam questions. This is a rather repetitive activity and they're fantastic at it. At my place we also have two official languages and we need to have exam assignments on both languages, guess who does the translation now (respecting LaTeX formatting, which previous machine translation tools typically wouldn't do).
- As an example of a one-off thing, the other day I had to edit a Word document which was password-protected. I asked ChatGPT how to unlock it and it not only answered, but actually did it for me (after 3 tries, but still, much faster than the time it would have taken for me to find out how to do it and then actually do it).
These are just some examples where they contribute (greatly) to my productivity at work. In daily life, I also ask them lots of questions.
1. Small coding tasks ("I want to do XYZ in Rust"), it has replaced stack overflow. Very convenient when writing code in a language I'm not super familiar with. 2. Help with English (traduction, proofreading...) 3. Learning something, like tech, I like interacting with it by asking questions, it's more engaging than just reading content.
I'd say nothing is game changing, but it's a nice productivity boost.
As a concrete example, I was recently playing with simulating the wave equation, and I wanted to try to use a higher-order approximation as I had never done that before. I'm quite rusty as I haven't done any numerical work since university some decades ago.
I still recalled how to deal with the Neumann boundary conditions when using the traditional lower-order approximation, but I was uncertain how to do it while keeping the higher-order approximation.
Searching for "higher-order neumann boundary conditions wave equation" or similar got me pages upon pages of irrelevant hits, most of them dealing with the traditional approximation scheme.
So I turned to ChatGPT which right away provided a decent answer[1], and along with a follow-up question gave me what I needed to implement it successfully.
[1]: https://chatgpt.com/share/67b4ab43-6128-8013-8e5a-3d13a74bf6...
I don't use integrated coding tools, so my workflow isn't super fast, but that's not what I'm really aiming for - more that I want to save my brain's energy from low level drudgy boilerplate or integration code, so I can focus it on the more important decisions and keep business-side context in my head.
It's been a huge help for me this way across multiple projects in multiple domains.
For example, in the beginning of this year, I completed this exercise where I wrote a lot about childhood, past experiences, strengths and weaknesses, goals and ambitions for the future, etc (https://selfauthoring.com) and then I uploaded all that to ChatGPT, asked it to be my therapist/coach, and then asked it to produce reports about myself, action plans, strategies, etc. Super interesting and useful.
By now ChatGPT has quite a bit of context from past conversations. Just the other day I used this prompt from someone else and got back super useful insights – "Can you share some extremely deep and profound insights about my psyche and mind that I would not otherwise be able to identify or see as well as some that I may not want to hear"
I did write 50 or more lines of instructions on what needs to be done and in what order.
ChatGPT gave me 5/6 (I asked for this) bash scripts totalling 300+ lines that seamlessly work together.
After reviewing, I asked it to change a few places.
If any human tried the same (except those rare bash Gods), it'd take many hours. I think it took me less than 30 minutes.
* when I use chatbots as search engines, I'm very quickly disappointed by obvious hallucinations
* I ended up disabling github copilot because it was just "auto-complete on steroids" at best, and "auto-complete on mushrooms" at worst
* I rarely have use cases where I have to "generate a plausible page of text that statistically looks like the internet" - usually, when I have to write about something, it's to put information that's in my head into other people head
* I'd love to have something that reads all my codebase and draws graphs, explain how things work, etc... But I tried aider/ollama, etc.. and nothing even starts making sense (is that an avenue to persevere in, though ?)
* At once, I tried to write in plain english a situation where a team has to do X tasks, in Y weeks, and I needed a table of who should be working on what for each week. I was impressed that LLMs were able to produce a table - the slight problem was that, of course, the table was completely wrong. Again, is it just bad prompting ?
It's an interesting problem when you don't know if you're just having a solution in search of a problem, or if you're missing something obvious about how to use a tool.
Also, all introductory texts about LLMs go into many details about how they're made (NNs and transformers and large corpuses and lots of electricity etc...) but "what you can do with it" looks like toy examples / simply not what I do."
So, what is the "start from here" about what it can really do ?