1,311 karma · joined December 10, 2024
I see two problems with LLMs & agents which wont be fixed possibly forever.
1) They dont have causal models. What they can do only is trial-and-error exploration which works quite well for many problems. But many other problems require a causal model.
2) Prompts lack precision, and programming languages and machine models were invented to solve this problem. English is great, but it is not a programming language.
Love this. I am frustrated by idiot software features everywhere, but am not triggered yet to punish them. AI automation is coming close however.
Sure, but what machine model do the public contract docs and RFCs target ? While the specific coding language is a detail, programming languages target a precise machine model. There is no machine model for english prompts and spec docs. So expect fuzziness - he-said-she-said bugs in your code - things that the LLM made up because the RFCs were not precise. And by the time you add precision to the specs, it has morphed into a new programming language with a machine model. In this usecase, since it is a port, the impact is mitigated because it is a clone of some existing functionality.
Well actually, spacex owned xai is now coding AI in C while earlier it was rust; I guess with the attitude that a language good enough to control rockets is fine for ai.
Reminds me of a colleague who rewrote the entire standard template library from scratch to demonstrate his uber-templating knowledge. He was the only one who could debug problems in that. The rest of us were ordinary c++ coders.
I have seen developers and managers having an attitude of arrogance just because they happen to code in a specific language. They also look down upon other developers just because they happen to code in some other language. Evangelism combined with arrogance is toxic.
Indeed. There is an irrational urge in some folks to become language-missionaries. Usually such folks have gained expertise in that specific language and want to protect and expand their turf. There is a wide-range of software usecases requiring a variety of tools and no one language fits all.
Amusing side-note. xai was all-in on rust for their ai-stack back in 2023. But now, spacex controlled xai is apparently coding ai in C - perhaps with the attitude that if a language is good enough to control rockets it is good for ai.
There are large areas of software where the quality of code is measured by the mental model of its human maintainers. For these areas, LLMs are a net negative because their use degrades the mental model of the human maintainers.
Love the metaphor. Planes are sophisticated machines capable of auto-piloting, but humans are still needed to ultimately pilot the beast.
google invests in anthropic and spacex - and shows appreciated values as earnings. Then it turns around and rents tpus to anthropic to show it as revenues. The main buyers and sellers for all of this are the hyperscalers, openai and anthropic.
It is a game of musical chairs while the party is still on.
Motivated reasoning in both camps.
There are folks incentivized by AI: engineers and managers working for AI related companies who justify their beliefs with selective facts. And then there are engineers who are threatened by AI and are extra-sensitive to slop.
The article is missing the point that once camps have made up their minds - no amount of analysis is going to change that.
Elon announces that spacex wrote its AI software in C. And now suddenly, C has become the new (old) kid on the block. Now we have folks saying, lets redo this in C as it gives you full power over the machine since we are 10x engineers. Earlier it was rust this or rust that. So, fads work both ways.
Of late the markets have become a casino. There is a large retail population that bets on options. Betting on IPOs is just another opportunity for such folks. By bending the rules Elon and others are trying to make it a more favorable bet for retail ensuring a near-term pop.
Kudos to S&P for standing up and sticking it to the man. And, woe to JP Morgan, Morgan Stanley etc for pushing overpriced paper, and on nasdaq etc for bending rules. This will be remembered later as the peak (my opinion).
So, tokens are used to produce sloppy code, and then this thing uses more tokens to fix vulnerabilities in the slop ? Whats not to like in this business model ? Similar to microsoft's. Create an OS which is vulnerable, and then enable business models for anti-virus software. Everyone wins.
More seriously, linters are turned off in ci because the amount of time spent chasing false-positives is prohibitive.
Use-it-or-lose-it is the evolutionary principle, both for cognitive and physical abilities.
subject/object dichotomy is a springboard to many schools of thought.
1) that subject emerges from objects - ie, anything has a material explanation, and everything is a machine.
2) that objects emerge out of a subject as a world model (platonic, descartes)
3) the subject and objects are one and the same representation of nature (spinoza)
4) subjects and objects emerge and disappear together (buddhist)
Anything reduced to computation is a fixed function from input to output, and is "dead" in the sense that it is unadaptable to its environment. Weights therefore is a dead machine.
Another view of this is that any closed system has unanswerable questions within it. Therefore, there is no system that can encompass everything. Hence weights being a closed system doesnt encompass everything.
Google and Amazon fund Anthropic which returns the favor with cloud purchases at these hyperscalers. So, google and amazon show increased earnings (via anthropic share markup) and increased cloud revenues via anthropic purchase. SpaceX didnt want to be left behind, so, it signed a deal with Anthropic.
Meanwhile capex at hyperscalers, VCs, PE etc is funding the party. Capex is not a concern to anybody as it doesnt appear on either revenues or earnings at the hyperscalers.
Downstream is partying from all the spending (server makers, chips, disk etc).
Whats not to like ! this is a perpetual money machine. Lets partay !
Agree, and I have the same worry, and so did the OP to the point that he considers canceling his AI subscription. Most of have to first get through the incidental friction to be able to arrive at expertise to address ambiguity.
AI tools have put friction where it should be - by eliminating incidental friction. By incidental friction I mean, things that were really not ambiguous, but were made so due to lack of access to resources.
As an example, if i needed to navigate, I used a paper map. There was friction in pulling out a map, planning a route etc. This took time. With digital mapping apps this sort of incidental friction is not there.
Real friction is inherent ambiguity. For example, what product does the market need ? By eliminating incidental friction, AI allows us to focus on the smallest hard-problem where there is real-friction.
Directly: 135M input tokens - $0.57 (134M cached)
Via OpenRouter 6M tokens - $0.81 (caching stats & inp/out not reported)
Caching is a huge win with using deepseek directly.
add deepseek v4 to it, and it will be close at 1/10 th the price. I use all three codex, claude, and deepseek, and they are close.
beautifully articulated, thanks @qubex
You misunderstand what mathematicians do. My favorite example is this: Birds can fly, but can they use their intuition to create a theory of flight, and then use that to build a plane ? This book by Bessis paints this really well [1].
[1] https://www.amazon.com/Mathematica-Secret-World-Intuition-Cu...