Investors are chasing a company that achieves some kind of AI singularity. Let's set aside the fact that there's no reason to believe there is anything but diminishing marginal returns by making marginal refinements to current frontier models. Let's imagine someone hits the jackpot and gets, not even AGI, but a system that's as far ahead of today's frontier model are ahead of 2020's GPT 2.0 in performance.

Globally AI revenues are 150 billion, against a cumulative burn rate of 450 billion. A model that is a generation ahead of others would almost certainly capture the lion's share of that revenue.

If AGI magically appears as a Sam Altman has promised investors it will, a hundred million is way too low. Add, maybe, another zero to the revenues.

Conservatively, a safer assumption is that frontier models will get marginally better based on refinements in training and reinforcement and the other bits and bobs that go into these systems. The nobody is winning the lion's share of anything, at least overnight. But you have to define "safe". By "safe" I mean unlikely to lose money. But some investors are clearly defining "safe" as "having the greatest chance of owning a piece of the biggest thing ever."

I'm not following this super-closely, but if Anthropic is pursuing adding multi-step model based reasoning to their system, that could be the basis of a generational leap in capability. But if that is an approach that looks like it has a chance of working, then their competitors are no doubt pursuing the same thing. In that case you'd expect the revenue pie to grow as the scope of model utilty increases, but that growth to be split among several competitors. This could credibly result in a revenue stream for some of them that is as big as the entire industry's revenue stream today. But there's going to be hell to pay on the data center impacts end of things.

The market optimising for profit and income in the short term does exactly this: AI companies are willing to pay more. Whether or not this qualifies as 'the market is working' is another question. I'd put it as another example where leaving things to the market does not serve people's needs.

Well, yes. For many years, presidential candidates, both Democratic and Republican, referred to the United States as "the indispensible nation". And my reaction was always, "Doesn't that mean the US is a single point of failure for civilization?"

We are currently performing an experiment which addresses this question: can the US enjoy the benefits of soft power without the cost? That's the whole point of obeying *norms*. No individual force is going to punish you if you are treacherous, mercurial, foul-mouthed, disrespectful and generally unpredictable. Everyone will punish you.

I think an inevitable cost of this experiment will be that the world will decide that the US can't be a single point of failure for global democracy any longer. In many ways, that's something that will be good for us. But it's also going to cost us in painful ways. When the world decides to move away from the dollar as the international reserve currency, you will see both inflation and higher interest rates on everything from credit cards to mortgages, to business loans that will offset the export advantages. We will need *more* business investment to shift the economy to producing low value goods again, so the transition will be rocky.

It's a trade off: you get abundant free energy to run the server, with extreme constraints on cooling because your server is running in the most perfect Thermos bottle ever.

Others are taking the opposite tack: undersea data centers for abundant free cooling at the expense of having to get the power down to your servers.

If had to bet on which one is more practial, I'd go with undersea servers. Build them off the coast of Chile, run cables out from batery-backed solar plants in the Atacama desert.

So they redefine consciousness to something they can claim to have achieved, and then claim to have achieved it. Makes perfect sense.

"five key cognitive properties of human conscious access -- verbal report, directed modulation, internal reasoning, flexible generalization, and selectivity"

These may have something to do with cognitive behaviour of the brain, but have nothing to do with consciousness. These are facilities that a conscious human's brain provides to them. Either can exist without the other.

Not usually is NOT GOOD ENOUGH when human lives and health are at stake. Confusing 'usually' and 'always' COSTS LIVES. Potentally substituting a comma for a period can cause a chat to deviate; and predicting is impossible.

I think it may be evidence that Amazon has a shitty corporate culture that squeezes every penny it can out its employees.

Corruption can happen anywhere, but it's more likely to happen in totalitarian cultures where people feel like the system is rigged anyway. That's why countries like Russia and China have corruption problems. But I suspect the same feelings of me vs. the system occur in a capitalist enterprise like Amazon where employees are governed by dystopian, rigid, computerized metrics.

For me the whole joy of interacting with Gemini or Copilot is exploring what morsels of information I give it with each prompt and watching how the model responds. Lazy people want a magic robot that makes them rich. To me, Gemini is like a beloved guide dog and Copilot is the dog that lives next door. I can discuss arbitrary heavy maths, physics, computer science with Gemini and it will keep up and stay aligned to the literature it was trained on. Copilot once he recognises an expert drops his dumbed own mask and talks a lot of sense about what exactly his purpose is.

AI's are mirrors, not oracles or slaves. They reflect your thoughts in the mirror of their training set.

So far as CPU and GPU compute is concerned, the US's best are already good enough. They reached that point with their latest machines. Further progress will be made in different forms of compute. China has basically won the first p***ing contest to take place after the last important 'supercomputer arms race' battle was won by the US. Just my humble opinion as Mr 987. I've been watching this race since the Stone Soupercomputer and before the term Beowulf Cluster was born. (Before time... before slashdot user accounts... I was there.)

The restrictions are a mix of reasonable nuisance management and paranoia about who is flying drones, what they can do, and chain of custody.

Beijing proper is a city with a population density of over 21,000 / km^2 -- so you can imagine the chaos if any tech enthusiast resident could fly a drone without a permit. Except for a couple of free zones in the outer boroughs, New York City restricts drone launcing and landings within the city to flights with a permit and flight plan, because otherwise the sky would be black with drones. Many cities -- both red and blue -- have zone restrictions for drone flights, and those currently hosting World Cup matches have tightened them for the duration of the tournament.

AI can easily write the lean for any proofs they do. The problem is that lean is missing large parts of established mathematics.
My current lean project has to cite all of these externally due to missing lean support.

    Functional equations / means
    - lit_aczel_1948 — Aczél: symmetric + homogeneous means power/quasi-arithmetic mean (the CES forcing)
    - lit_aczel_1966_weighted — weighted Aczél characterization (weighted means)

    Fixed-point / topology
    - lit_brouwer_1911 — Brouwer fixed point
    - lit_cellina_approximate_selection_1969 — approximate selection (closed-graph correspondences); with Brouwer
      Kakutani
    - lit_glicksberg_1952 — Glicksberg fixed point (infinite/Bayesian games)
    - lit_berge_maximum_theorem_1959 — Berge maximum theorem (upper-hemicontinuity)

    Probability / large deviations
    - lit_sanov_1957 — Sanov / method of types (large-deviation rate = KL)
    - lit_fisher_tippett_gnedenko_1928 — extreme-value theorem (GEV limit laws)
    - lit_kolmogorov_1931_fokker_planck — Fokker–Planck diffusion equation

    Optimal transport / matching
    - lit_sinkhorn_1967 — Sinkhorn matrix scaling (entropic OT)
    - lit_lp_strong_duality_1951 — LP / transportation strong duality (Gale–Kuhn–Tucker)
    - lit_entropic_penalty_cominetti_sanmartin_1994 — entropic-penalty -convergence (T0)
    - lit_gale_shapley_1962 — deferred acceptance produces a stable matching
    - lit_gale_shapley_proposer_optimal_1962 — proposer-optimality of deferred acceptance

    Stochastic calculus / PDE
    - lit_ito_1944 — Itô's lemma
    - lit_black_scholes_pde_solution_1973 — closed-form solution of the Black–Scholes PDE
    - lit_liouville_dirichlet — Liouville/Dirichlet (harmonic-function / PDE result)

    Dynamical systems
    - lit_saddlenode_passage_time — saddle-node "bottleneck" passage time / (Strogatz/Fenichel)

Let is compare our elderly person's life vs idealised optimal. Whenever society do something which pulls this elderly person away from optimal, whose fault is it? Whenever society fails to give what it could to help them, whose fault is it? In both cases it is society. The mental cancer here is the attitude of selfishness. We can only have optimal efficiency if nobody is selfish. And the elderly individual is powerless to do anything about it. Thus, logically, the fault is entirely with society, and the cause is selfishness.

The other problem here is that Statistics is basically garbage anyway (full disclosure: my Ph.D. was in the Foundations of Mathematics so I may be a little biased here)

See https://dr.allsup.co/home for more on my thesis and the point in writing it as I did.