acituan

> why wouldn't capital drive the value of inference to the absolute bare minimum?

Capital can drive the cost of inference down but the value is determined by whoever wants to pay for that inference. Therefore this question is not answered: how valuable really is these llm inferences and how much of a supply of such inference does the world really need.

I understand your point about misattribution but it cuts both ways. How about when a company is better than competitors because they executed better because they had a superior organizational culture. Or not successful and this is due to poor culture.

YC sets the prime examples. It is never product at the expense of who the team is and in what proven way they have worked together and plan to execute at scale.

Why would you assume culture is immaterial? And to make this less emotional let’s take the micro scale; don’t you think the culture of doing engineering doesn’t affect outcomes team to team within the same company, or company to company within the same country or even country to country within the same company?

Here is where I got that impression:

> once the optimizations became sophisticated enough

Either way I am not trying to litigate here. Feel free to correct me if your position was softer.

I don't know why you'd think your analogy wasn't clear in the first place. But your analogy can't support you on the assertion that optimizations will be sophisticated and reliable enough to completely forget about the programming language underneath.

If you have any first principles thinking on why this is more likely than not, I am all ears. My epistemic bet is that it is not going to happen, or somehow if we end up there the language we will have to use to instruct them is not going to be different than any other high level programming language that the point will be moot.

First problem is turning engineers into accountability sinks. This was a problem before LLMs too, but now a much bigger and structural problem with democratization of the capacity to produce plausible looking dumb code. You will be forced to underwrite more and more of that, and expected to absorb the downsides.

The root cause is the second problem; short of formal verification you can never exhaustively prove that your code works. You can demonstrate and automate that demonstration for a sensible subset of inputs and states and hope for the state of the world approximately staying that way (spoiler: it won't). This is why 100% test coverage in most cases is something bad. This is why sensible is the key operative attitude, which LLM suck at right now.

The root cause of that one is the third problem; your job is to solve a business problem. If your code is not helping the business problem, it actually is not working in the literal sense of the work. It is an artifact that does a thing, but it is not doing work. And since you're downstream of all the self-contradicting, ever changing requirements in a biased framing of a chaotic world, you can never prove or demonstrate that your code solves a business problem and that is the end state.

> compilers only got better and better

At no point compilers produced stochastic output. The intent user expressed was translated down with a much much higher fidelity, repeatability and explainability. Most important of all, it completely removed the need for the developer to meddle with that output. If anything it became a verification tool for the developer‘s own input.

If LLMs are that good, I dare you skip the programming language and have it code in machine directly next time. And it is exactly how it is going to feel like if we treat them as valuable as compilers.

> What got me past that point was short bursts at BPMs way past my comfort zone and building synchrony _after_ I stumbled upon more efficient motions that scaled.

This is actually pretty close to what Stetina says. I just probably didn’t do a good job expressing it.

You’re oscillating above and below the comfort zone and that iteration like you say affords insights from both sides, and eventually the threshold grows.

Great suggestion of a video, I’ll check it out.

Funny how this exactly applies to instrument playing. Unearned speed only begets sloppiness. The only way to go past a certain velocity is to do meticulous metronome work from a perfectly manageable pace and build up with intention and synchrony. And even then it is not a linear increase, you will need to slow back down to integrate every now and then. (Stetina's "Speed Mechanics for Lead Guitar"; 8 bpm up, 4 bpm down)

Language is not humanness either; it is a disembodied artifact of our extended cognition, it is a way of transferring the contents of our consciousness to others or to ourselves over time. This is precisely what LLMs piggyback on and therefore are exceedingly good at simulating, which is why the accuracy of "is this human" tools are stuck at %60-70's (%50 is a coin flip), and are going to be bounded for a foreseeable future.

And I am sorry to be negative but there is so much bad cognitive science in this article that I couldn't take the product seriously.

> LLMs can be scaled almost arbitrarily in ways biological brains cannot: more parameters, more training compute, more depth.

- Capacity of raw compute is irrelevant without mentioning the complexity of computation task at hand. LLM's can scale - not infinitely - but they solve for O(n^2) tasks. It is also amiss to think human compute = a singular human's head. Language itself is both a tool and protocol of distributed compute among humans. You borrow a lot of your symbolic preprocessing from culture! Like said, this is exactly what LLM's piggyback on.

> We are constantly hit with a large, continuous stream of sensory input, but we cannot process or store more than a very small part of it.

- This is called relevance, and we are so frigging good at it! The fact that machine has to deal with a lot more unprioritized data in a relatively flat O(n^2) problem formulation is a shortcoming, not a feature. Visual cortex is such an opinionated accelerator of processing all that massive data that only the relevant bits need to make to your consciousness. And this architecture was trained for hundreds of millions of years, over trillions of experiment arms - that were in parallel experimenting on everything else too.

> Humans often have to act quickly. Deliberation is slow, so many decisions rely on fast, heuristic processing. In many situations (danger, social interaction, physical movement), waiting for more evidence simply isn't an option.

- Again a lot of this equivocates conscious processing to entire cognition. Anyone who plays sports or music knows to respect the implicit, embodied cognition that goes on to achieve complex motor tasks. We are yet to see a non-massively-fast-forwarded household robot do a mundane kitchen cleaning task, and go play table tennis with the same motor "cortex". Motor planning and articulation is a fantastically complex computation; just because it doesn't make it to our consciousness or instrumented exclusively through language doesn't mean it is not.

> Human thinking works in a slow, step-by-step way. We pay attention to only a few things at a time, and our memory is limited.

- Thinking, Fast and Slow by Kahneman is a fantastic way of getting into how much more complex the mechanism is.

The key point here is as limited in their recall, how good humans are at relevance, because it matters, because it is existential. Therefore when you are using a tool to extend your recall, it is important to see its limitations. Google search having indexed billions of pages is not a feature if it can't bring the top results well. If it gets the capability to sell me whatever it brought up was relevant, that still doesn't mean the results are actually relevant. And this is exactly the degradation of relevance we are seeing in our culture.

I don't care if the language terminal is a human or a machine, if the human was convinced by the low relevance crap of the machine it just a legitimacy laundering scheme. Therefore this is not a tech problem, it is a problem of culture; we need to be simultaneously cultivating epistemic humility, including quitting the Cartesian tyranny of worshipping explicit verbal cognition that is assumed to be locked up in a brain; we have to accept that we are also embodied and social beings that depend on a lot of distributed compute to solve for agency.

We know from the era of data the power of JOIN. Bring in two different data sources about a thing and you could produce an insight neither of them could have provided alone.

LLMs can be thought as one big stochastic JOIN. The new insight capabilities - thanks to their massive recall - is there. The problem is the stochasticity. They can retrieve stuff from the depths and slap them together but in these use cases we have no clue how relevant their inner ranking results or intermediary representations were. Even with the best read of user intent they can only simulate relevance, not really compute it in a grounded and groundable way.

So I take such automatic insight generation tasks with a massive grain of salt. Their simulation is amusing and feels relevant but so does a fortune teller doing a mostly cold read with some facts sprinkled in.

> → I solve problems faster by finding similar past situations → I make better decisions by accessing forgotten context → I see patterns that were invisible when scattered across time

All of which makes me skeptical of this claim. I have no doubt they feel productive but it might just as well be a part of that simulation, with all the biases, blind spots etc originating from the machine. Which could be worse than not having used the tool. Not having augmented recall is OK, forgetting things are OK - because memory is not a passive reservoir of data but an active reranker of relevance.

LLMs can’t be the final source of insight and wisdom, they are at best sophists, or as Terrence Tao put it more kindly, a mere source of cleverness. In this, they can just as well augment our self-deception capacity, maybe even more than counterbalancing them.

Exercise: whatever amusing insight a machine produces for you, ask for a very strong counter to it. You might be equally amused.

> I'm always surprised how many 'logical' tech people shy away from simple determinism, given how obvious a deterministic universe becomes the more time you spend in computer science, and seem to insist there's some sort of metaphysical influence out there somewhere we'll never understand. There's not.

You might be conflating determinism with causality. Determinism is a metaphysical stance too because it asserts absence of free will.

Regardless of the philosophical nuance between the two, you are implicitly taking the vantage point of "god" or Laplace's Demon: infinite knowledge AND infinite computability based on that knowledge.

Tech people ought to know that we can't compute our way out of combinatorial explosion. That we can't even solve for a simple 8x8 game called chess algorithmically. We are bound with framing choices and therefore our models will never be a lossless, unbiased compression of reality. Asserting otherwise is a metaphysical stance, implicitly claiming human agency can sum up to a "godlike", totalizing compute.

In sum, models will never be sophisticated enough, claiming otherwise has always ended up being a form of totalitarianism, willful assertion one's favorite "framing", which inflicted a lot of pain in the past. What we need is computational humility. One good thing about tech interviews that it teaches people resource complexity of computation.

My bad, skipped a chain of thought there. Since medicare pays less than private insurance, hospitals can and do shift costs (which in reality is "opportunity cost of profit") to the latter, which pushes to private premiums up. Regardless, this is a minor effect. Very little of the inflation is justified with productivity gains, as you said it is a very inefficient healthcare system. US prices clock 2x-4x of comparable OPEC peers, admin percent is higher etc.

Employers paid for healthcare in 1970s too, and even for higher percentages of the workforce. If there is a premium inflation surpassed the CPI, that is still inflation, not real growth. If there’s an inflation problem in delivering a temporally comparable service, that is not a “real wage” item for the employee [1]. So what the nominal figure today shouldn’t be relevant.

I agree it shouldn’t be an employer item too, but whatever employers lose on premiums, they get more on an overall stickier and cheaper labor supply.

[1] one could argue the productivity of healthcare increased, and the data indeed supports this with the overall life expectancy increase from 70s to now mid 70s plus quality of life treatments. But again most of the spend is actually on the tail end at this age group, which raises the workers’ premium without delivering the benefit. Therefore not much structural gain for the actual working age employee.

> what are fair ways to extract value from citizens for the shared value of the state?

The right question is who benefits the most from state’s services. For example if a whole lot of security, legislative or admin services go to protecting the capital, then those who has the most capital need to chip in the most.

> redistribution is usually that “more” people reach a higher standard of living, then adding taxes and friction to processes like automation may conflict with that goal

This is basically a 50 year old trickle down argument. But real wages have not increased in comparison to gdp since 70s, so nothing trickled down. We are demonstratedly bad at sharing what we have achieved together, no reason to believe more tech will magically get better treatment than that.

Besides redistribution is not about shifting the curve up, but making it flatter - see gini coefficient.

> the core benefit of automation, which is to delete non-needed work, make things cheaper, and make the value creator richer.

Except the era of classical capitalism and inventor’s profit is over, since 70s it is rentiers unreciprocated extraction on top of purported value people didn’t necessarily ask for or need in the first place. Likewise most people aren’t dying for AI automation, and not even for structural threats; it is not even proven that it will provide a net total productivity gain when the hype cools down, despite being shoved down people’s throats.

Let’s not kid ourselves, there is little concern for real value creation but a capture-the-flag on a gigantic data-moated compute monopoly. Whatever democratic means enabled proper taxation would have already prevented this type of speculative berserk, failures of which I assure you will be socialized.

So friction = societal consent, internalizing externalized costs, revealing what is actually value versus monopolist’s rent. It is healthy for the society, it is healthy for capitalism.