Certainly the “unsupervised agent” workflows are getting a lot of attention right now, but they require a specific set of circumstances to be effective:
- clear validation loop (eg. Compile the kernel, here is gcc that does so correctly)
- ai enabled tooling (mcp / cli tool that will lint, test and provide feedback immediately)
- oversight to prevent sgents going off the rails (open area of research)
- an unlimited token budget
That means that most people can't use unsupervised agents.
Not that they dont work; Most people have simply not got an environment and task that is appropriate.
By comparison, anyone with cursor or claude can immediately start using this approach, or their own variant on it.
It does not require fancy tooling.
It does not require an arcane agent framework.
It works generally well across models.
This is one of those few genunie pieces of good practical advice for people getting into AI coding.
Simple. Obviously works once you start using it. No external dependencies. BYO tools to help with it, no “buy my AI startup xxx to help”. No “star my github so I can a job at $AI corp too”.
Using nano banana does not require arcane prompt engineering.
People who have not learnt image prompt engineering probably didn't miss anything.
The irony of prompt engineering is that models are good at generating prompts.
Future tools will almost certainly simply “improve” you naive prompt before passing it to the model.
Claude already does this for code. Id be amazed if nano banana doesnt.
People who invested in learning prompt engineering probably picked up useful skills for building ai tools but not for using next gen ai tools other people make.
Its not wasted effort; its just increasingly irrelevant to people doing day-to-day BAU work.
If the api prevents you from passing a raw prompt to the model, prompt engineering at that level isnt just unnecessary; its irrelevant. Your prompt will be transformed into an unknown internal prompt before hitting the model.
Because running a parallel process is often difficult. In most cases, the question becomes:
So, how exactly is my app/whatever supposed to spin up a parallel process in the OS and then talk to it over IPC? How do you shut it down when the 'host' process dies?
Not vaguely. Not hand wave "just launch it". How exactly do you do it?
How do you do it in environments where that capability (spawning arbitrary processes) is limited? eg. mobile.
How do you package it so that you distribute it in parallel? Will it conflict with other applications that do the same thing?
When you look at, for example, a jupyter kernel, it is already a host process launched and managed by jupyter-lab or whatever, which talks via network chatter.
So now each kernel process has to manage another process, which it talks to via IPC?
...
Certainly, there are no obvious performance reasons to avoid IPC, but I think there are use cases where having the compiler embedded makes more sense.
AI is hard; edge cases are hard. AI sucks at edge cases.
Between AI for cars and AI for software the long tail of edge cases that have to be catered for is different, yes.
...but I'm sure the same will apply for AI for art (e.g. hands), and AI for (insert domain here).
Obviously no analogy is perfect, but I think you have to really make an effort to look away from reality not to see the glaringly obvious parallels in cars, art, programming, problem solving, robots, etc. where machine learning models struggle with edge cases.
Does the tooling they used matter? no, not at all.
...but if they've claimed to solve the 'edge case problem', they've done something really interesting. If not, they haven't.
So, don't claim to have done something really interesting if you haven't.
You can say "I've been using AI to build a blah blah blah. It's great!" and that's perfectly ok.
You have to go out of your way to say "I've been using an AI to build blah blah blah and I haven't written any of it, it's all generated by AI". <-- kinda attention seeking.
"no lines of code directly written" really? Why did you mention that? You got the AI to write your software for you? That sounds cool! Let's talk! Are you an AI consultant by any chance? (yes, they are). ...but.
No. You didn't. You really didn't. I'm completely happy to call people out for doing that; its not unfair at all.
> My estimate is that 80-90% of the code was written by AI
Nice! It is entirely reasonable both to do that and to be excited about it.
…buuut, if that’s what you’re doing, you should say so.
Not:
“no lines of code directly written, just directing the AI”
Because those (gluing together AI code by hand and having the agent do everything) are different things, and one of them is much much MUCH harder to get right than the other one.
That last 10-15%. Self driving cars are the same story right?
"no lines of code directly written, just directing the AI"
/skeptical face.
Without fail, every. single. person. I've met who says that, actually means "except for the code that I write", or "except for how I link the code it build together by hand".
If you are 50kloc in to a large complex project that you have literally written none of, and have, eg. used cursor to generate the code without any assistance... well, you should start a startup.
...because, that's what devin was supposed to be, and it was enormously and famously terrible at it.
So that would be either a) terribly exciting, or b) hyperbole.
Oh, wait. I guess it's a supported use case. I uh... take it back... I guess...
> If you don’t want to use a dedicated CRM platform or start from a template, you can use a no-code platform like Notion to build a knowledge base or team homepage and modify it to fit your CRM needs.
Yeah, I guess some people will think that's a good idea.
> it took me several hours to get llama.cpp working as a server
Mm... Running a llama.cpp server is annoying; which model to use? Is it in the right format? What should I set `ngl` to? However, perhaps it would be fairer and more accurate to say that installing llama.cpp and installing ollama have slightly different effort levels (one taking about 3 minutes to clone and run `make` and the other taking about 20 seconds to download).
Once you have them installed, just typing: `ollama run llama3` is quite convenient, compared to finding the right arguments for the llama.cpp `server`.
Sensible defaults. Installs llama.cpp. Downloads the model for you. Runs the server for you. Nice.
> it took me 2 minutes to get ollama working
So, you know, I think its broadly speaking a fair sentiment; even if it probably isn't quite true.
...
However, when you look at it from that perspective, some things stand out:
- ollama is basically just a wrapper around llama.cpp
- ollama doesn't let you do all the things llama.cpp does
- ollama offers absolutely zero way, or even the hint of a suggestion of a way to move from using ollama to using llama.cpp if you need anything more.
Here's some interesting questions:
- Why can't I just run llama.cpp's server with the defaults from ollama?
- Why can't I get a simple dump of the 'sensible' defaults from ollama that it uses?
- Why can't I get a simple dump of the GGUF (or whatever) model file ollama uses?
- Why isn't 'a list of sensible defaults' just a github repository with download link and a list of params to use?
- Who's paying for the enormous cost of hosting all those ollama model files and converting them into usable formats?
The project is convenient, and if you need an easy way to get started, absolutely use it.
...but, I guess, I recommend you learn how to use llama.cpp itself at some point, because most free things are only free while someone else is paying for them.
Consider this:
If ollama's free hosted models were no longer free and you had to manually find and download your own model files, would you still use it? Could you still use it?
If not... maybe, don't base your business / anything important around it.
It's a SaaS with an open source client, and you're using the free plan.
They failed to provide any examples of facts with regard to Devin.
This is like arguing that it’s not fair to critique people claiming to have made superconductors because “some people said they are really superconductors” but no one can share samples with anyone for some reason.
A reasonable counter argument would be:
> Here is evidence of Devin actually doing things.
How, other than the available evidence was anyone supposed to evaluate Devin?
There is a broad opportunity for the developers to respond to this, but they haven’t.
Why is that?
It is because he’s right.
Regardless of what Devin can do that video was deceptive and misleading. There no two ways about it.
> Also the result is somewhat counterintuitive. We know that by low level of understanding, if we ask a student a hard question and he tried many times, the most accurate answer is often not the most popular one but a single answer.
This is a bad analogy.
Here’s what is actually happening with no “common sense but wrong” understanding of it:
- You have a set of probabilities per token.
- You randomize them.
This is not a “bad student being asked multiple times” it is a system with randomized probabilities, creating a probability distribution.
If you want to see what a probability distribution looks like (eg. An electron cloud) then sampling only once is the wrong way to do it.
You basically have two distributions; the first one is the LLM, the second one is the shape generated by adding the random factor in the temperature.
This allows you to escape the “local maxima” encoded in the LLM distribution to find highly probable solutions that are outside the sample space of the “zero temperature”.
If you want a better analogy, look up at the night sky full of stars. Draw circle in the sky; that’s the LLM distribution.
The result from a zero temperature will be the brightest point in that circle.
When you push the temperature up, you blur the sky randomly. Some points become brighter, some dimmer, but the radius of the circle increases.
If there is a very bright point outside the sample circle 10x brighter than the brightest point inside it then repeated random samples will repeatedly find it.
It makes perfect sense that an expanded probability distribution sampled repeatedly could find a “good average solution” if that solution is significantly better than the best “zero temp” solution.
This is the same reason we have 'temp' at all; by widening the solution space probability distribution, you can find better maxima. Turns out, sampling multiple times lets you have more chances to find better maxima.
This is more like "well that seems obviously like a good idea" than "somewhat counterintuitive"; it's just slow and expensive to do it.
You can also adjust the probability distribution by other existing methods, obviously, what's surprising here is not that it works, but that it seem to work so well; probably (and I note they did not try this in their paper), a multi-sample + voting on the output from other methods would also be highly effective.
I mean, SDXL is great. Until you’ve had a chance to actually use this model, isn’t calling it out for some imagined offence that may or may not exist seems like you’re drinking some Kool-aid rather than responding to something based in concrete actual reality.
You get access to it… and it does the google thing and puts people of colour in every frame? Sure, complain away.
You get access to it, you can’t even generate pictures of girls? Sure. Burn the house down.
…you haven’t even seen it and you’re already bitching about it?
Come on… give them a chance. Judge what it is when you see it not what you imagine it is before you’ve even had a chance to try it out…
Lots of models, free, multiple sizes, hot damn. This is cool stuff. Be a bit grateful for the work they’re doing.
…and even if sucks, it’s open. If it’s not what you want, you can retune it.
> Long ago he wrote a book entitled The Design of Everyday Things. It's well worth the read. Within those pages he stated the following rule of thumb: “If you think something is clever and sophisticated beware -- it is probably self-indulgence.”
> You also asked me "why...". To the extent that I have not answered that above, I'll simply turn the question around and point out that it is probably for the same reason that your video was solely focussed on the amplification of performance to the strident denigration of every other concern. To a performance hammer, everything looks like a nail. ;-)
Overall a very amicable and interesting read.
It’s so easy to preach high level architecture without specifics, but also so very easy to cherry-pick a specific example where generic advice doesn’t apply.
I think it’s interesting that there is an almost fundamental disconnect between “easy to understand” and “fast and efficient” …and I’m absolutely 100% with Uncle Bob that bounded contexts for complex code is the solution.
This is the approach rust takes with unsafe code and it has proved to be an extremely effective principle.
Am I cynical in thinking that if you vote yes, it’s a permanent record on your Twitter account, which he can then use later to mete out anti-musk retaliation punishments?
I mean, it’s not like he’s used the platform to do punish anyone who opposes him or his views before.
The 'proko' channel on youtube has some discussions on this topic from people who know what they're talking about, rather just anecdotal evidence from a sample size of 1.
The diffusion art models make it very clear that these technologies go from toys to replacing people’s jobs very very quickly.
Short of a (not impossible) breakdown in civilization, the current rate of change suggests that 5-10 years is probably unrealistic.
More like 2-4 years, it will be technically possible to be reasonably high level software engineering.
It’s already possible to chain high level “how would I do x?” Planning to “write code to do y…” code generation.
It’s very very obvious that technically it’s going to be possible to do a great deal of mechanical work like refactoring and adding trivial features automatically.
Companies do not need 50 engineers doing 100x productivity.
They need like, 5.
What remains to be seen is if governments and legal systems allow that scale of displacement of human labor to occur or not.
> As it continued to grow, I additionally trademarked the name “Gitea” in order to protect the project’s brand. (More on the trademark later.)
...
> I have also transferred both the domains and trademarked name to Gitea Ltd. so that they are no longer personally owned by me and will remain indefinitely with the Gitea project.
(no further mention of trademarks)
So, I think the take away here is as a response to the core demands of the 'open letter' being:
- A non-profit organisation owned by the Gitea community is created.
- The Gitea trademark and domains are transferred to the non-profit.
- The name of the company is changed to avoid any confusion with the non-profit.
The answer is bluntly; no.
With regard to the concerns raised about use use of a DAO, the response is (as previously):
> One of the options we have been considering includes a decentralized autonomous organization (DAO). ... The DAO management model would also not mean the creation of a gitcoin or crypto token.
So, basically no acknowledgement that using a DAO is fundamentally indistinguishable from using crypto.
What, I ask, is the point of having laws and rules if you can just ignore the ones you don't like?
Its just a name, who cares?
Not me.
…but, if you break the law, you break the law. Not maybe maybe who cares, its not me being water boarded, I dont care…
If you break the law. You break the law.
Otherwise, who gives a duck what congress says?
Just fire them all and crown Trump King of America.
I’m being facetious. …but maybe its more of a big deal than you superficially pretend it is.
It’s just another case of the administration blatantly breaking the rules.
…so, you know. If youre ok with no laws or rules, I guess its fine.
Seems a bit chaotic to me. I prefer my governing body to be… marginally bound by some kind of responsibilty to something or someone.