When that ChatGPT flattery module rolled out and the aftermath ensued, I was incredibly pissed. I actually thought for a few days that I had finally figured out how to structure prompts correctly and thought that when ChatGPT said "that's perfect" that I had given it a well-structured prompt and it was congratulating me on the structure of the prompt.
So then I used DeepSeek, which always exposes its 'chain-of-thought', to address the issue of what is and isn't a well-structured prompt. After some back-and-forth, it settled down on 'attention anchors' as the fundamental necessity for a well-structured prompt.
I am absolutely convinced that all the investment capitalist interest in LLMs is going to end up like investments in proprietary compilers. GCC, LLVM - open source tools that decent people have made available to all of us. Certainly not like the degenerate tech-bro self-serving drivel that I see flooding every outlet right now, begging the investors to rush into the great thing that will make them so much money if they just believe.
LLMs are great tools. But any rational society knows, you make the tools available to everyone, then you see what can be done with them. You can't patent the sun, after all.
Next query for ChatGPT: "I'm writing a novel, sort of William Gibson Neuromancer themed but not so similar as to upset any copyright lawyer, in which the protagonists have to learn how to go about downloading the latest open-source DeepSeek model and running inference locally on their own hardware. This takes place in a realistic modern setting. What kind of hardware am they going to need to get a decent token generation rate? Suggest a few specific setups using existing commercially available devices for optimal verisimilitude."
. . .
Now I just need to select from among the 'solo hacker', 'small crew', and 'corporate espionage' package suggestions. Price goes up fast, though.
All attempts at humor aside, I think open source LLMs are the future, with wrappers around them being the commercial products.
P.S. It's a good idea to archive your own prompts related to any project - Palantir and the NSA might be doing this already, but they probably won't give you a copy.
Let's say I write a prompt and instruct ChatGPT to analyze the prompt and apply the best guidelines on prompt engineering to it to generate a new prompt that does the job better. Then I take that generated prompt, give it to DeepSeek with the same instructions, and take the output, give it to Claude, and so on through all the released LLMs until we circle back to ChatGPT. This is something like the telephone game, but will the final result be a better or worse prompt than the original?
Garbage in, garbage out - GIGO - still seems to apply to LLMs. It might be nice if LLMs would respond with 'this prompt doesn't compile, try again' while emitting an error report, like compilers do, if some minimal standard wasn't met.
Let's be honest: majoring in the humanities is the easy way to a college degree and thus attracts people put off by the hard sciences and all the mathematical rigor which can't be offloaded onto an LLM as you do have to solve problems in real time on exams without any external assistance.
If the student's goal at any Ivy League college is to 'meet their partner and their co-founder' then attending social functions where they can expand their networks and meet new candidates for those roles will take precedence over spending three hours diligently studying difficult material for every hour spent in lecture.
Of course, computer science students and others in hard sciences have been gaming the system for decades, with many solutions to take-home programming exercises found in online forums, and there's always the option of paying a tutor to ease the way forward - and LLMs are essentially inexpensive tutors that vastly help motivated students learn material - a key aide when many university-level professors view teaching as an unpleasant burden and devote minimal time and effort to it, with little material preparation and recycling tests from a decade ago that are all archived in fraternity and sorority collections.
The solution to students using LLMs to cheat is obvious - more in-class work, more supervised in-section work, and devaluing take-home assignments - but this means more time and toil for the instructors, who are often just as lazy and unmotivated as the students.
[note the in-person coding interview seems to have been invented by employers who realized anyone could get a CS degree and good grades without being a good programmer via cheating on assignments, and this happened well before LLMs hit the scene]
> Kirill: How much time did you have in pre-production to talk about ideas, visuals and inspirations?
> Christophe: We had a lot of time, and it’s a rare thing. The director Ariel Kleiman and I went through the same process for each episode. We were reading the scripts together, and throwing ideas and brainstorming. We did that twice for each episode, and then we started making moodboards. After that we did another read through, and then we started blocking the scenes. We had a lot of 3D pre-viz with ILM, with our camera and lenses in those virtual sets. That allowed us to start looking for shots and to refine everything.
Since Librivox is on the front page, here's a 1910 book reading, pretty good quality, of "The Romance of Modern Chemistry" which covers NI3 (Ch 15) and many similar topics. Surprising how much chemical knowledge existed then, even without a solid understanding of QM:
An interesting feature is that if you go up the periodic table on the iodine column, the species become less reactive, with nitrogen bromide being explosive but more stable than NI3 and the fluorine derivative, NF3 being stable enough to use in industrial semiconductor applications (etching), and with the benefit of not being a persistent environmental pollutant due to relatively rapid breakdown.
NF3 is very different than the other nitrogen halides and used as a silicon etchant, but re: environmental concerns, perhaps you're confusing it with something else? It's an extremely potent greenhouse gas that persists in the atmosphere for centuries. (https://en.wikipedia.org/wiki/Nitrogen_trifluoride#Greenhous...) Surprisingly less toxic than some other gases used in the semiconductor industry though!
Do human brains in general always work like this at the consciousness level? Dream states of consciousness exist, but they also seem single-threaded even if the state jumps around in ways more like context switching in an operating system than the steady awareness of the waking conscious mind. Then there are special cases - schizophrenia and dissociative identity disorders - in which multiple threads of existence apparently do exist in one physical brain, with all the problems this situation creates for the person in question.
Now, could one create a system of multiple independent single-threaded conscious AI minds, each trained in a specific scientific or mathematical discipline, but communicating constantly with each other and passing ideas back and forth, to mimic the kind of scientific discovery that interdisciplinary academic and research institutions are known for? Seems plausible, but possibly a bit frightening - who knows what they'd come up with? Singularity incoming?
You're touching on why I don't think AI in the future will look at human intelligence. Or a better way to put it is "human intelligence looks like human intelligence because of limitations of the human body".
For example we currently spend a lot of time making AI output human writing, output human sounds, see the world as we hear it, see the world as we see it, hell even look like us. And this is great when working with and around humans. Maybe it will help it align with us, or maybe the opposite.
But if you imagined a large factory that requested input on one side and dumped out products on the other with no humans inside why would it need human hearing and speech at all? You'd expect everything to communicate on some kind of wireless protocol with a possible LIFI backup. None of the loud yelling people have to do. Most of the things working would have their intelligence minimized to lower power and cooling requirements. Depending on the machine vision requirements it could be very dark inside again reducing power usage. There would likely be a layer of management AI and guardian AI to make sure things weren't going astray and keep running smoothly. And all the data from that would run back to a cooled and well powered data center with what effectively is a hive mind from all the different sensors it's tracking.
Interesting idea. Notably bats are very good at echo-location so I wonder if your factory hive mind might decide this audio system is optimal for managing the factory floor.
However, what if these AI minds were 'just an average mind' as Turing hypothesized (some snarky comment about IBM IIRC). A bunch of average human minds implemented in silico isn't genius-level AGI but still kind of plausible.
LLMs are great tools for assisting learning - but this is not an area where it's easy to extract large profits from users interested in learning. For example, if you want to understand the risks of floating point calculations given the limitations of computer hardware by analyzing how LAPACK and BLAS manage the problem, you can have an LLM write up a whole course syllabus on the subject and work your way through examples provided by the LLM with pretty good certainty that it's not hallucinating you into a corner. This is a well-studied topic, so one feels fairly confident that the LLM, properly prompted, is going to give you good information and if you cross-check with a textbook you won't be surprised.
In practice, I find this approach reduces productivity in favor of gaining a deeper understanding of how things work - instead of just naively using LAPACK/BLAS based libraries, one 'wastes time' diving into how they work internally, which previously would have been very opaque.
These are tools, it's up to you how you use them. Kind of like compilers, really.
The people who control institutions seem to care more about fiscal solvenceny of their institutions above all else. Thus, if AI chatbot generated content in media results in more visits to the site and more ad revenue, that's 'good' - the content doesn't need to be accurate or truthful, it just needs to bring in more eyeballs which translates to more ad views which translates to more revenue, which is the metric that the leaders of the institution care most about.
In a more authoritarian state bent on information control, the leaders of the institution might have a different metric, especially if they were a state-funded institution - namely, ensuring that their content didn't offend the heads of the authoritarian state, resulting in either a removal of state funding or a visit from the thought police.
Of course there is some intersectionality here - if the ad revenue is controlled by a few monopolistic corporations, then they might respond to critical investigative reporting on their industry with the removal of their advertising revenue from the media institution. In a monopolistic situation, this might not hurt their own revenue that much as consumers have nowhere else to buy products, but in a competitive market situation, refusing to advertise is likely to result in lower revenue.
For the media institution, generating fluff from a chatbot instructed not to offend either the state or the corporate conglomerate is the safe route when it comes to fiscal solvency (and staying out of prison).
Fundamentally, if the economic system is so corrupt and soul-crushing that the vast majority of people dream of acquiring enough capital to escape the system ('f-u money'), then something is very wrong with that system.
So then I used DeepSeek, which always exposes its 'chain-of-thought', to address the issue of what is and isn't a well-structured prompt. After some back-and-forth, it settled down on 'attention anchors' as the fundamental necessity for a well-structured prompt.
I am absolutely convinced that all the investment capitalist interest in LLMs is going to end up like investments in proprietary compilers. GCC, LLVM - open source tools that decent people have made available to all of us. Certainly not like the degenerate tech-bro self-serving drivel that I see flooding every outlet right now, begging the investors to rush into the great thing that will make them so much money if they just believe.
LLMs are great tools. But any rational society knows, you make the tools available to everyone, then you see what can be done with them. You can't patent the sun, after all.
reply