If the parts you need from a larger third party library are small and simple, that makes sense. But I don't think the general idea of reusing battle tested code across many projects is obsolete because of AI.
I think the approach should probably move towards semi-tailored systems.
Maybe you don't need all of Postgres - so provide 2,3,4,5,n versions. As long as every version is covered by (possibly shared) tests, then you can release what works, ignore the rest (bugs and all).
Exactly. And over time a project can drift from its original purpose, so you end up building around a bunch of stuff you aren't using anyway, but you pay cost/risk for having it present.
I don't see the author complaining about it, instead he thought it was an interesting story. So I guess he already did that.
I have my business email read first by an LLM, I imagine someone who works at a tech-oriented website does too so those emails are basically filtered into a pile anyway.
Put very simply, that is how they're supposed to work. They are supposed to result in the most likely result. Same reason they, unless made not to, output '17' when prompted to "pick a number between 1 and 30", because that is the most likely result from training; but that means that all LLMs output 17 except for forced temperature or reasoning in RL making them run a python command with random.randint(0,31), or that the prompt is recognised and something else is supposed to be chosen than what is statistically the most likely text response.
Therefore asking a group of people would result in more varied outcomes of something visual, but from the textual training data the direction is statistically the most likely so is what is output, that sky is the most likely, time is the most likely, cloud coverage, etc. You're prompting with as little as possible so it fills in the blanks with what as likely as possible. When you ask Alex from the North Pole they're going to fill in the blank with a different kind of weather than Bob from the Equator.
That does mean that an LLM itself is in essence not as creative as a human, but then it is a tool to be used by a human doing more than oneshotting "Generate an SVG of a pelican riding a bicycle" and expecting creativity from the tool rather than the user. If the user is lacking in creativity they can prompt an LLM to get wacky scenarios to change it up a little.
>They are supposed to result in the most likely result.
That's called greedy decoding and is never used in practice. Everybody's using random sampling, so it's the range of most likely results. This issue is all about that range being way smaller than the training data implies.
Assembly doesn't add anything here. You'd still need to implement the IO, portability, safety, etc etc. You'd end up with the same thing so you might as well use Go from the get go. Assembly is useful for dedicated SIMD like in dav1d.
It's part of their job. If they can't do that right they certainly can't do the rest of their job right. Just like it's OK to get queazy if your doctor is functionally illiterate or your lifeguard can't swim. Also things that match average people in some areas.
I've been "forced" to use AI in my work since this year, but I only started experimenting with it with my own ChatGPT subscription recently. I'm able to afford $20/mo and I'm fine paying it. The average user meta is targeting wants a free assistant to replace Siri, which is what muse is.
Continuing my thought, talking to Muse and ChatGPT feel fundamentally different. The single thread chat Muse offers encourages you to talk to it like a person imo. But that's probably just me.
> Security from websites, installable software, physical access?
All those would be important, but don't let the perfect be the enemy of the good. We could start from the Vulnerability Of The Year: supply chain attacks.
Being able to visit a web page, open a PDF or download and inspect a git repo without the web page, PDF or git repo's being able to completely pwn the entire machine including the bootloader and the firmware for all the hardware subsystems.
Desktop Linux is very bad at that compared to ChromeOS and MacOS, but some distros are much better than others. For a few decades now, Fedora has been one of the better ones.
I use it in the car to talk to because it's built-in, mostly for navigation via voice. It seems to be an old or quantised model, because it seems a few years old, and it has strict rate limits unlike Gemini, where it really just shuts off if you talk to it a few times in one day, but it's still fun to show passengers who are still new to LLMs.
reply