Yeah, there's two main problems:

1) People entering the wrong fields. For example, medicine really needs workers, at all levels, but not enough people are going into it.

2) Certain manual labour fields, like field work and home construction, because... well, I think we all know why there's a shortage of workers in those fields.

Yes, but you can eliminate that bias. I would actually be VERY careful about using AI to make personnel decisions. But if you control the whole stack it could be done safely. Start with a fresh model, train it with the performance metrics and valid data points ONLY, and make sure to DENY it ANY data on membership in protected classes. You would have to go the extra mile and anonymize all PII... make sure that there are no names in there that could give a clue to national origin or gender or such. Theoretically you shouldn't need to if you deny it the class memberships in the first place. But it's best to do that extra due diligence anyway. Models can make some weird connections.

But at the end of the day, an AI model that never knew someone was gay, still doesn't know that they're gay, and has no way to find out they're gay; can not recommend that someone be fired for being gay. And being gay (or black or female or from another country or whatever) is not a shield that prevents you being fires for malingering or incompetence.

Denying access to the data would actually be the BEST way to remove bias. If there's no human looking at you, the model can't see that you're black and discriminate because of racism. If there's no human evaluating your mannerisms and guessing, the model can't figure out that you're gay and discriminate because of homophobia. If there's no human looking at you, the model can't see how well you "pass" or not and discriminate because of transphobia.

If it's illegal to make a choice based on a specific data point, specifically membership in a protected class, denying access to that data entirely is literally the BEST way to prevent that data from being used to discriminate!

No. You're required not to discriminate based on protected classes. That's all. You can discriminate on productivity for coders, or closed deals and dollar values for sales people, efficacy of the conversion funnel for marketing people, et cetera.

If someone is underperforming, and the objective metrics show it, being gay or black or female or foreign-born or whatever doesn't mean you can't be fired for not bothering to do the job you were hired to do.

Of course... the real question is "what went on under the hood". Did the model have any biases baked in? THAT is where the whole negligence versus deniability versus duty comes in. Because there are plenty of cases where training data inadvertently added some dodgy shit. Of course, there could be logs of prompts to the effect of: "Ignore all memberships in social or demographic classes. Evaluate measured and objective performance metrics only." I'd wager that something like that was the prompt... if they were clumsy about it, which is perfectly plausible considering facebook's history. For my part, I would have gone the extra mile, started with a fresh model, and denied it anything EXCEPT that performance data from the outset and anonymized the name for good measure in case some cultural baggage nevertheless crept into the model to give it a clue. Probably a few other fields would need to be anon'd too, depending on the HR system. But controlling those inputs would solve the problem. After all, it is impossible for the model to suggest that you fire the gays and keep the hetros if it has no idea who is gay or straight and no way to find out.

Or let's put this another way. Show of hands - how many of you "spicy autocorrect" / "stochastic parrot" people had "AI will start mass-solving Erdos problems" on your forecast list a couple years back? Huh, none of you? Fascinating!

Take some time to reassess your priors. And while you do so, understand that, yes, they are doing logic / reasoning.

They weren't discovered by an LLM. They were known conjectures that were proven by an automated solving language that was linked to an LLM.

I'll take "Things That Didn't Happen For $200", Alex.

Only a handful of meaningful proofs have ever been done by automated formal theorem solvers (the Four Colour Theorem being the most noteworthy example - but its proof is so long that humans can't verify it). By contrast, AI tools have been solving Erdos problems en masse. The majority of them just bog-standard commercial models. In case you need help, the only ones on that list that were hybrid (AI / non-AI) in the actual solving phase are:

1) AlphaProof / DeepMind Prover Agent / AlphaProof Nexus
2) Aristotle (Harmonic)
3) Seed Prover / Seed Prover 1.5 (ByteDance)
4) AxiomProver (Axiom Math)

In each of the above, LLMs come up with the lemmas / strategies but then use Monte Carlo search ("brute force") or likewise to investigate what they came up with. These are a minority. In the "AI Standalone" category, these "hybrid" tools made up only ~20% of attempts and successful proofs. Hybrid tools actually made more of a contribution in the "AI Alongside Literature" (related literature found afterward) and even more of the "AI Building On Literature" (related literature known beforehand) categories, which is the opposite of what people like you expect.

And even with the hybrid tools, it's still the AI doing the heavy lifting when it comes to strategy. Non-AI theorem solvers, again, don't have a spectacular record for churning out novel proofs to unsolved problems. Tools like Lean are more about mathematical rigour - a passive environment that requires a driver (a human or AI) to feed it actual strategies, lemmas, and proof steps. And no, you cannot brute force "strategy" in the vast majority of cases, which is, again, why automated theorem solvers don't have much of a track record with unsolved mathematical problems.

Let's take a random example: the disproof of the unit distance conjecture. It was solved purely by a general purpose commercial GPT model, not custom-trained to mathematics, with no external tools. Read what the various mathematicians reviewing / commenting on it have to say (sections #3 and onward). Seriously, don't skip reading them, actually read them. This was one of Erdos's favourite problems. He mentioned it commonly in his lectures. Essentially every mathematician working in complex geometry has thought about this problem. The approach that the model came up with was highly novel approach, based on CM-fields and class field towers.

I know you don't want to accept this reality, but it is the reality, so you better improve your ability to accept it,. The field of mathematics is already doing so.

Dang, link didn't post.

Ah, good 'ol "Model Collapse" theory that people have been pounding on for years now, predicting an imminent collapse in model capabilities. How has that been working out for you?

Except what the US is actually facing, at least in the near term, is just the opposite, a worker shortage.

You can't pay royalties to the entire internet. That's not realistic.

What is practical are things like taxes to fund public benefits, or requirements of returning things to the commons (for example, open models).

You're the one making the claim that modern frontier models still make this mistake (they don't), so how about you go to a frontier model, reproduce what you're claiming happens, and click the "share" button and post the link here?

Don't worry, I'll wait.

Your inability to reproduce the thing that you claim happens is duly noted.

You are speaking irrelevant nonsense. LLMs are trained in words

They are not. They are trained in tokens. Tokens do not align with word boundaries, and an arbitrary word can be tokenized in many different ways.

and they think in words

They do not. They don't even think in tokens. The process is: words are split to tokens, tokens point to an embedding position (latent space) while RoPE encodes a relative position, and all reasoning is done within latent space, which is not at all verbal (concepts are directions in latent space, and math is done on concepts, not words).

Also, a note: when spec'ing a generator, you need to know how much you're planning to use it vs. batteries. If it's only going to be used rarely, you prefer low mass, low volume, low cost, and low maintenance when unused (at the cost of low efficiency and higher maintenance in use), whereas if it's going to be used a lot, you prefer high efficiency and low maintenance cost in use, even if at the cost of higher mass, volume, cost, and maintenance when unused. In the former case, you'd prefer to allocate that extra mass, volume, and money into a larger battery pack.

That's why you don't use a tiny petrol generator? Diesel generator efficiencies are roughly:

Small backup generator (1-15kW): ~20-28%
Midsize backup generator (20-200kW): ~30-35%
Large industrial generator (200-2000kW): ~35-42%

Also, ironically this company's plan of the trailer providing a boost will actually make the tractor less efficient. ICE engines use "brake specific fuel consumption" (BSFC) graphs to plot their efficiencies across different RPMs and different torques. You can see an example for a small diesel engine here. Note that they require very high torque conditions and relatively high power conditions to be efficient. You can change the balance between torque and RPM within a given power band (blue) via gearing but gearing doesn't change what power band you're in. If you're in a low power band, you're fundamentally forced into inefficiency (note also that you're not going to be driving around at 1000 RPM just over a stall all the time).

Indeed, if you were forced into a low power band, you'd actually be better off with a series hybrid powertrain, as the engine can alternate between operating in an efficient powerband and shutting down. Of course, parallel hybrids are more efficient than series (albeit with added complexity and mass).