HN user
You don't have to have health insurance at all,
There are (or were) tax time penalties for failing to have healthcare coverage. Possible USA laws have changed recently.
What gp wanted to say is that models are now so smart and useful that even if they managed to be EVEN MORE smart and useful, you wouldn't even notice it.
If benchmarks across the board keep trending up and you still don't notice a difference, that's not evidence the model stopped improving. More likely your tasks aren't hard enough to expose the gains, or the model has passed the point where you're able to judge it.
You can only tell a good answer from a great one up to your own ceiling. Once the model clears that, both look the same to you, and the extra capability is real whether or not you can see it.
"--speculative-config",
Regarding that last option: speculation helps max concurrency when it replaces many memory-expensive serial decode rounds with fewer verifier rounds, and the proposer is cheap enough. It hurts when you are already compute-saturated or the acceptance rate is too low. Good idea to benchmark a workload with and without speculative decoding.
specific type of manipulable gullibility
A congregation trained to say “amen” on cue.
Not a bad idea, but how do we prevent this from creating incentives to engineer similar situations in the future?
Sound like Singapore's system rewards competence + stability, and that makes it hard for opposition to scale, because the incumbents can absorb popular ideas while keeping institutional advantages.
That’s demoralizing if you’re doing party-building. But the flip side is: it means citizens do have leverage, just not always in the form of “replace the government.”
Money is less about personal consumption and more about a voting system for physical reality. When a company holds billions in IOUs, they are holding the power to decide what happens next. That capital allows them to command where the next million tons of aluminum go, which problems engineers solve, and where new infrastructure is built.
Even if they never spend that wealth on luxury, they use it to direct the flow of human effort and raw materials. Giving it away for free would mean surrendering their remote control over global resources. At this scale, it is not about wanting more stuff. It is about the ability to organize the world. Whether those most efficient at accumulating capital should hold such concentrated power remains the central tension between growth and equality.
FYI: Newer LLM hosting APIs offer control over amount of "thinking" (as well as length of reply) -- some by token count others by an enum (high low, medium, etc.).
I hear you on intergenerational stuff. I just don’t think “public education is a scam” fits what most kids actually receive.
Kids are not only getting classroom time. They inherit a whole baseline that previous taxpayers built: safer streets, clean water, courts that mostly function, vaccines, roads, libraries, stable money, and the accumulated tech and culture that makes modern jobs even possible. That bundle is huge, and it starts paying out long before anyone is old enough to “owe” anything.
Also, adults are not literally trapped. People can move, downshift, opt out of a lot, or choose different communities. Most don’t, even when they complain loudly, and to me that’s a pretty strong signal the deal is at least somewhat reasonable. Not perfect. Not fair for everyone. But not a cartoon pyramid scheme either.
If there’s a real fight worth having, it’s making the burdens and benefits less lopsided across generations, not pretending the whole social investment in kids is fake.
I do not believe there exists a way to safely use LLMs in scientific processes.
What about giving the LLM a narrowly scoped role as a hostile reviewer, while your job is to strengthen the write-up to address any valid objections it raises, plus any hallucinations or confusions it introduces? That’s similar to fuzz testing software to see what breaks or where the reasoning crashes.
Used this way, the model isn’t a source of truth or a decision-maker. It’s a stress test for your argument and your clarity. Obviously it shouldn’t be the only check you do, but it can still be a useful tool in the broader validation process.
There is no single definition for all scientists. However if you define free will as choices that are completely free of deterministic or even statistically deterministic causes that science could in principle predict, then most scientists would say: no, that kind of free will probably doesn’t exist.
closed lid would flex slightly in your backpack, and press keys on the keyboard,
FYI: Over time, this repeated pressure + rubbing (especially with dust or grit in between) can leave permanent key-shaped marks or “ghosts” on the screen. Thin laptops and bags that are tightly packed or bulging make this a lot more likely, since there’s less rigidity and more pressure on the lid.
I mostly agree with your intuition, but I’d phrase it a bit differently.
Temperature 0 does not inherently improve “quality”. It just means you always pick the highest probability token at each step, so if you run the same prompt n times you will essentially get the same answer every time. That is great for predictability and some tasks like strict data extraction or boilerplate code, but “highest probability” is not always “best” for every task.
If you use a higher temperature and sample multiple times, you get a set of diverse answers. You can then combine them, for example by taking the most common answer, cross checking details, or using one sample to critique another. This kind of self-ensemble can actually reduce hallucinations and boost accuracy for reasoning or open ended questions. In that sense, somewhat counterintuitively, always using temperature 0 can lead to lower quality results if you care about that ensemble style robustness.
One small technical nit: even with temperature 0, decoding on a GPU is not guaranteed to be bit identical every run. Large numbers of floating point ops in parallel can change the order of additions and multiplications, and floating point arithmetic is not associative. Different kernel schedules or thread interleavings can give tiny numeric differences that sometimes shift an argmax choice. To make it fully deterministic you often have to disable some GPU optimizations or run on CPU only, which has a performance cost.
I agree that consumption is the point of production in the long run, but that does not mean consumption is neutral in terms of tradeoffs. When someone consumes, the underlying real resources are gone and cannot be used for other purposes. That is the core of the broken window point: activity can rise while net wealth falls if resources are used in less productive ways than the alternatives. Saying consumption is not "taking from society" skips over the opportunity cost of what could have been built with the same labor and capital.
On who controls investment, it is fair to worry about political influence, but it does not follow that highly concentrated private wealth is mostly idle or socially useless. Large fortunes are usually claims on productive assets that employ people and produce goods and services. Capital markets already involve broad and diverse mechanisms like index funds, pensions, and institutional investors that pool the savings of millions of non wealthy people and allocate them based on expected returns. That is not perfect, but it is not simply a handful of billionaires directing everything according to whim.
The equation of exchange point also overstates the role of velocity in judging what is good for the economy. A dollar that flows into an asset is not disappearing from the real economy. It is funding someone else who is selling equity or debt and who will use that for wages, R&D, or capital spending. Lower velocity at the cash register can be consistent with higher long run output if more of today’s income is channeled into projects that raise productivity tomorrow. High velocity tied to fragile consumption and low investment can look good in the short term but leave people poorer over time.
One of the big advantages humans have is that we are really good at working around constraints rather than just fighting over them. Health care, education, and even things like prenatal screening to reduce serious genetic disease all expand people’s potential and quality of life. Over time, that means more people who are emotionally stable, educated, and healthy, which is exactly what makes “higher quality” partners possible in the first place.
When we treat everything as a zero sum status game, especially when it comes to other human beings, we tend to slide into ugly, destructive dynamics (up to and including wars) where everyone burns resources just to keep their relative place. Cooperation and innovation, on the other hand, are how we turn a fixed pie into a growing one.
So yes, relative position matters in some contexts, but the whole reason our species has gotten this far is that we are not limited to fighting over scraps. We can change the size and shape of the pie together.
Capitalism ties decision power to how good you are at accumulating wealth. Other systems give decision power through birth, status, or bureaucracy. But so far none have matched capitalism at growing total wealth over time- just look at “communist” China adopting capitalist tools to get rich, and now struggling with the inequality that comes with it.
You’re slipping into a version of the broken window fallacy here. Consumption isn’t automatically good for society just because it “creates activity.” When you eat an apple, the apple - and all the labor and resources behind it - are gone. Same with luxury goods: the more resource-intensive they are, the more real wealth they permanently use up compared with simpler alternatives.
What actually increases long-run prosperity isn’t consumption itself, but efficient use of resources and investment - choices that expand the stock of tools, knowledge, and productive capacity in the future. New wealth is created when resources are not immediately consumed, but are instead used to boost productivity.
That’s why wealthy people who don’t spend all their wealth aren’t “hoarding” in the economic sense. Their capital is usually invested - financing factories, startups, research, tools, and infrastructure that generate more output. And when they die with wealth still invested, the state captures a chunk through estate and capital-gains taxes.
A better way to think about wealth is as decision-making power over how resources are allocated - toward current consumption (including luxuries) or toward future production (investment). Capitalism tends to concentrate that decision power in the hands of those who are best at growing capital, which raises total prosperity but also increases inequality.
Consumption uses up wealth; investment grows it. People like Musk have a lot of wealth because they’ve been good at growing it. We should absolutely guard against wealth hijacking politics - but it would be shortsighted to treat their continued investment as a net negative for society.
Faced with subtle lies you may prefer the obvious ones.
One thing I keep wondering, though, is whether “life” is tied more to the particular chemistry and environment it uses or to its patterns (the abstract information structure that can, in principle, be re-instantiated on different substrates).
If it’s the patterns that matter, do you think it’s actually impossible for those patterns to be transmitted across interstellar distances? Just like a cup of ocean water is packed with DNA, it’s at least conceivable that what we call “cosmic background noise” could, in principle, hide extremely compressed life-patterns that only an advanced civilization could recognize and reconstruct back into something we’d meaningfully call “alive.” And of course, the more efficiently you code that information, the more it statistically has to look like random noise.
Not saying this is likely -- just that if the essence of life is informational rather than chemical, "traveling" could look very different for any life that is suitably advanced.
evolution by natural selection only applies when the adaptation in question affects survivability before reproducing
Anything that happens to you after reproductive age does not get affected by natural selection because the selection pressure of reproduction is gone.
You are mostly correct but must also consider traits that affect the odds children will fail to reproduce. People can for example be genetically predisposed to depression or impulsive anger or substance abuse or ... any of which can impact the survival of their children thus selection pressure does not entirely disappear after a child is born.
Our entire societal system is based on increasing revenue (due to inflation).
Yes that is capitalism however if inflation cuts value of money in half and in the same time your revenue doubles, did you actually double your revenue? Do you even need to change your service or product to justify raising prices when the currency is being devalued? For both these questions there is a strong case that the answer is no.
I get the worry. AFAIK most of the current capex is going into scalable parallel compute, memory, and networking. That stack is pretty model agnostic, similar to how all that dot com fiber was not tied to one protocol. If transformers stall, the hardware is still useful for whatever comes next.
On reasoning, I see LLMs and classic algorithms as complements. LLMs do robust manifold following and associative inference. Traditional programs do brittle rule following with guarantees. The promising path looks like a synthesis where models use tools, call code, and drive search and planning methods such as MCTS, the way AlphaGo did. Think agentic systems that can read, write, execute, and verify.
LLMs are strongest where the problem is language. Language co evolved with cognition as a way to model the world, not just to chat. We already use languages to describe circuits, specify algorithms, and even generate other languages. That makes LLMs very handy for specification, coordination, and explanation.
LLMs can also statistically simulate algorithms, which is useful for having them think about these algorithms. But when you actually need the algorithm, it is most efficient to run the real thing in software or on purpose built hardware. Let the model write the code, compose the tools, and verify the output, rather than pretending to be a CPU.
To me the risk is not that LLMs are a dead end, but that people who do not understand them have unreasonable expectations. Real progress looks like building systems that use language to invent and implement better tools and route work to the right place. If a paper lands tomorrow that shows pure next token prediction is not enough for formal reasoning, that would be an example of misunderstanding LLMs, not a stop sign. We already saw something similar when Minsky and Papert highlighted that single layer perceptrons could not represent XOR, and the field later moved past that with multilayer networks. Hopefully we remember that and learn the right lesson this time.
Human reasoning is, in practice, much closer to statistical association than to brittle rule-following. The kind of strict, formal deduction we teach in logic courses is a special, slow mode we invoke mainly when we’re trying to check or communicate something, not the default way our minds actually operate.
Everyday reasoning is full of heuristics, analogies, and pattern matches: we jump to conclusions, then backfill justification afterward. Psychologists call this “post hoc rationalization,” and there’s plenty of evidence that people form beliefs first and then search for logical scaffolding to support them. In fact, that’s how we manage to think fluidly at all; the world is too noisy and underspecified for purely deductive inference to function outside of controlled systems.
Even mathematicians, our best examples of deliberate, formal thinkers, often work this way. Many major proofs have been discovered intuitively and later found to contain errors that didn’t actually invalidate the final result. The insight was right, even if the intermediate steps were shaky. When the details get repaired, the overall structure stands. That’s very much like an LLM producing a chain of reasoning tokens that might include small logical missteps yet still landing on the correct conclusion: the “thinking” process is not literal step-by-step deduction, but a guided traversal through a manifold of associations shaped by prior experience (or training data, in the model’s case).
So if an LLM doesn’t collapse under contradictions, that’s not necessarily a bug; it may reflect the same resilience we see in human reasoning. Our minds aren’t brittle theorem provers; they’re pattern-recognition engines that trade strict logical consistency for generalization and robustness. In that sense, the fuzziness is the strength.
So how is it therefore not actual thinking?
Many consider "thinking" something only animals can do, and they are uncomfortable with the idea that animals are biological machines or that life, consciousness, and thinking are fundamentally machine processes.
When an LLM generates chain-of-thought tokens, what we might casually call “thinking,” it fills its context window with a sequence of tokens that improves its ability to answer correctly.
This “thinking” process is not rigid deduction like in a symbolic rule system; it is more like an associative walk through a high-dimensional manifold shaped by training. The walk is partly stochastic (depending on temperature, sampling strategy, and similar factors) yet remarkably robust.
Even when you manually introduce logical errors into a chain-of-thought trace, the model’s overall accuracy usually remains better than if it had produced no reasoning tokens at all. Unlike a strict forward- or backward-chaining proof system, the LLM’s reasoning relies on statistical association rather than brittle rule-following. In a way, that fuzziness is its strength because it generalizes instead of collapsing under contradiction.
The challenge is that human values aren’t static - they’ve evolved alongside our intelligence. As our cognitive and technological capabilities grow (for example, through AI), our values will likely continue to change as well. What’s unsettling about creating a superintelligent system is that we can’t predict what it -- or even we -- will come to define as “good.”
Access to immense intelligence and power could elevate humanity to extraordinary heights -- or it could lead to outcomes we can no longer recognize or control. That uncertainty is what makes superintelligence both a potential blessing and a profound existential risk.
100%!
I'm not sure what would motivate an artificial intelligence.
Those who give it orders hence your concern about how AI will be used as a tool is spot on.
AI trained and built to gather information, reason about it, and act on its conclusions is not too different from what animals / humans do - and even brilliant minds can fall into self-destructive patterns like nihilism, depression, or despair.
So even if there’s no “malfunction”, a feedback loop of constant analysis and reflection could still lead to unpredictable - and potentially catastrophic - outcomes. In a way, the Fermi Paradox might hint at this: it is possible that very intelligent systems, biological or artificial, tend to self-destruct once they reach a certain level of awareness.
when you have some supremely intelligent agent acting on the world, even a small misalignment may end up in catastrophe
Why not frame this as challenge for AI? When the intelligence gap between a fully aligned system and a not-yet-aligned one becomes very large, control naturally becomes difficult.
However, recursive improvement — where alignment mechanisms improve alongside intelligence itself — might prevent that gap from widening too much. In other words, perhaps the key is ensuring that alignment scales recursively with capability.
Whatever it takes to add object permanence and long-term memory assimilation to LLMs may not be so easy to run on your 4090 at home.
Today yes but extrapolate GPU/NPU/CPU improvement by a decade.
for any problem there is an ideally sized model that when trained will produce the lowest error rate.
You studied ML before discovery of "double descent"?