Well, in theory theory, temperature 0 doesn't really exist. Mathematically, as lim temperature->0, the distribution gets spikier and spikier, the most likely sample goes to almost-but-not-quite infinity and the rest go to almost-but-not-quite 0. In practice, temperature=0 is literally a separate branch of an if statement that just picks the most common sample (using the actual formula that works for non-zero values would cause a zero division).

However, due to things such as batching and even different kinds of floating point imprecisions for different algorithm implementations, the probability distribution itself often differs run-by-run, so what you sample from it also differs.

I've been studying AI for 20 years. What really needs to be added to this statement is:

"An alarming number of people don't understand that LLMs work via purely stochastic processes - and so does human thinking. People do NOT arrive at the same conclusion if merely the weather's different. Worse: with human thinking not only do most people not think this is real, a subset of people will actively fight the idea. Of course, depending on the weather"

using low temperature is more deterministic, but the cost is the model becomes "dumber"

You don't even need temperature 0, just make a random seed for the sampler part of the input and then its deterministic as a function of the input.

But running autoregressive models at temp=0 tends to expose pathological behavior, because the training process produces a function with a lot of gain so its prone to feedback on its own noise.

The implementation does not often differ run by run.

I don't know for sure, but I would be surprised if it was illegal in my particular US state. You might be able to argue the AI has inherent biases that introduce illegal discrimination in the hiring process, but my understanding is winning I case like that would be very difficult, especially since most employers are very cagey about their hiring process and why they mades a decision.

Show someone a list of resumes with an "applicant score*" and they'll naturally ignore the ones with a low ranking

*scores are generated with AI, mistakes may be made, use only as a guide and verify results

If the first 50 people who apply are all bots, why are you reading resumes in order of submission?

1. Give them some easy leetcode questions. Nothing that a competent programmer would have any problem with.

2. If they pass, ask for a deposit of like $20. Shouldn't be an issue for people who are actually serious.

3. Do more simple leetcode questions but this time on zoom so you can tell if they are using AI. If they pass that they get the deposit back.

(Yeah I know there are real-time interview cheat AI programs but based on what I've seen on demos of them it's super obvious when they're being used.)

Probably not practical but just a thought!

It does exist very much, even if you go to pure math. Look at the softmax function and take the limit as T->0. It becomes a dirac-delta function. I.e. in a discrete setting (like for LLMs with a finite set of output tokens), probability P becomes one for argmax and 0 for everything else. Only in coding practice it is easer to implement T=0 as a simple if check that directly chooses argmax instead of calculating the limit of some function that includes 1/T quotients. But setting T to zero is in both, theory and practice, turning the usual probability function into greedy sampling.

But theoretically, the output of every LLM is seed-driven (or could be if you wrote the software to isolate it) just like any computer software. It's just none of the software written (even llama.cpp AFAIK) chooses to support stable-seeding due to the changes in stuff like CPU/Vulkan/CUDA/Metal differences making it difficult to make consistent.

They could though! Hopefully one day someone implements it into the mainstream LLM-engine software and it gets exposed in the APIs serving the models. It'd do a lot to show folks the "internals" of these models.

It is a common misconception, but it is not true even in principle. If I have 2 or more logits which are equal to the maximum of my logits, I will sample uniformly random from them with any temperature, even zero. Sampling from softmax([1, 0, 1]) is still stochastic at temperature 0, because the limit is to sample uniformly from the first or the last element.

Anyway: "GPUs don't do deterministic matrix multiplications" is the biggest source of randomness in LLMs. GPUs put the associativity of the sums in matrix multiplications in arbitrary order, and this has a huge impact on the logits coming out of the neural network.

If you train two different LLMs and replace what data they "see" in batch n, that doesn't affect the data they see in batch n+1, or any further batches. In LLMs, you can introduce "noise" into the training process, but that noise doesn't really compound.

Humans learn from experience, not from data, and their experiences at age n shape what experiences they seek (and hence train on) at age n+1. A small amount of "noise" injected into their "training", let's say hearing a group of friends discuss a movie while their identical tween goes to the bathroom, can compound into them watching that movie, which can compound into them forming an identity around that genre, and so on, until they're two completely different people, trained on completely different "data mixtures".

*According to our proprietary, undisclosed, non-deterministic metric, which may or may not be Math.random

Gates that reduce resume flow-through are only useful if their reduction is correlated with quality. Otherwise they're just dragging out your hiring process or unnecessarily causing you to ultimately lower your hiring bars.

https://stackoverflow.com/questions/16833100/why-does-the-mo...

Provided:

* If it’s MoE we are talking about, that the duplicate inputs are for the whole batch (yes, your batch neighbours can impact your choice of experts. Blergh.)

* Your kernels are deterministic

* There’s no system wide effort switch that responds to, e.g. work load across the cluster (for a thinking model)

Upshot:

Temperature 0 is not deterministic in probably any existing cloud infra, but it could be for edge inference pretty reliably.

To your quibble on 0.1 being more deterministic - I think it’s a pretty fair summary - we’re going to sample much more from the ‘temp 0’ answer at 0.1 than we would at temp 0.9, no?

Ah yes, the much revered cosmological fairness constraint.

The volume is infeasible to review everyone for quality, even at an hour scale. The conclusion and solution is inevitable, though I wish it were different. 35% is actually really good if you’re not coming in through a referral.

The current reality is <1% and the person reviewing you is exhausted.

I'm saying this as somebody who most of the time has some side project going on.

https://en.wikipedia.org/wiki/Secretary_problem

Far worse would be different humans having the same weights.

What I'd really love is an actual number for a "human hallucination rate". How often will a random human

1) claim something that is wrong

2) defend the wrong claim and/or logic even when the problem is pointed out to them

(and this of course outside of the usual topics. In politics? I don't care. In religion? Don't care (well, maybe a bit more than politics). Let's say in physics or popular logic or something like that)

After that RL step, you want to stick to the RL distribution, and so keep a temperature of 1.0. Other temperatures will drive the model out-of-distribution.

That is why the sampling step for agents or thinking LLMs are usually kept at a temperature of 1.0.

To contextualize this insight in your post and basically just repeat what you are saying: The mistake is not using a non-deterministic system. The mistake could be, in some sense, using it too little. Re-evaluating the same resume 5 times and seeing a high variance in scores is a more useful signal than evaluating it once.

It's totally fine to filter out resumes in a completely random, content-independent way. Grabbing the fourth resume down in the pile and offering them the job is a perfectly fair albeit stupid way to make a hiring decision. However, AIs are very, very good at capturing biases, and it would not at all surprise me if an AI told to filter resumes is going to end up filtering with some biases for things that you definitely do not want to filter on, like the name of the candidate. And it might be that everybody resume that claims it fixed a typo in a major open source project gets a pass, but resumes that only list their own projects get rejected 60% of the time, so you're losing more good candidates than bad.

Credentialing helps maintain a quality floor. Does this person have basic employable skill? Nothing more. It actually doesn’t help you identify levels of talent and skill which is a universal hiring problem.

We do have a credential - a CS degree. And you can see it is a mixed signal. Employers can choose of their own free will to take risks on employees that do have this credential, or not.

Mandating by law that you must have a CS degree doesn’t seem to help our field as we famously have high performers across the spectrum of formal education.

Also, it doesn't pick up certifications or awards. I tried some PRs people are suggesting with enhancements (https://github.com/Zem-0/hiring-agent), it helps, but overall their ATS is hugely biased towards people with large GitHub contributions to OSS.

Chickens coming home to roost.

Really depends on the program. In my undergrad program there were some very smart CS students who got great grades that really struggled with the programming. Smart and capable people can be bad at programming and lack many qualities that make for a good hire.

(A more charitable interpretation would be that aforementioned CTO was making a joke that didn't land.)

In pure math, it does not always do that. It becomes a dirac-delta comb with equal weight on every maximum. There can be more than 1 maximum. Setting the temperature to zero turns into greedy sampling, but greedy sampling is not necessarily deterministic as you can have multiple equally optimal options.

I don't understand the distinction you're drawing. A Dirac delta function is a "simple if check".

https://x.com/FireworksAI_HQ/status/2069873437217276015

But this isn't a fundamental property of LLMs, it's just an implementation detail. It's pretty obvious that if you evaluate the matrix multiplications correctly and deterministically sample from the highest-probability outputs, you will have a deterministic LLM.

That’s user-controlled too, not an inherent property of GPUs:

https://docs.pytorch.org/docs/2.12/generated/torch.use_deter...

We expect computers to be consistent despite running programs that are not designed to be consistent.

This despite the fact that we have lots of experience of programs running on computers that produces wildly inconsistent outputs.

But for some reason some people choose to assume LLMs should act like a calculator instead of any of those programs.

Due to acting like an irrational gambling machine, I agree it can have unwanted indirect discrimination effect in general. But it will probably not differentiate "on the grounds of religion or belief, disability, age or sexual orientation". It is possible, but that would take a lot of work for the lawyers to prove to the court.

I believe the more interesting part is that the EU AI Act (still not in force in this regard until 2 December 2027). This will be clearly a high-risk AI system: "AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates".

Which does not mean prohibited, but it could later turn out that LLMs will be excluded from being used in high-risk AI use cases (falling under article 6 with no exemptions).

Considering that none of the standards are published yet, I have absolultely no idea how they will ensure compliance with the following parts of Article 10 when using LLMs for such tasks: "(f) examination in view of possible biases that are likely to affect the health and safety of persons, have a negative impact on fundamental rights or lead to discrimination prohibited under Union law, especially where data outputs influence inputs for future operations; (g) appropriate measures to detect, prevent and mitigate possible biases identified according to point (f)"

I don't think that's technically possible to do so with LLMs in general at the moment, even with the full cooperation of the model providers. Maybe you can do some meaningful audits for smaller models. But the EU AI Act may end up excluding all the generic "using-LLM-but-not-entirely-sure-why" vibe coded approaches from high-risk use cases (in Annex III). Which would make sense.

https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng

Which means there's a good chance this is somehow correlated in one way or another to race/gender/other protected classes in the US, just by the math of everything being correlated to everything.

Which means this is one good lawsuit away from being illegal in the US as well. It doesn't even necessarily have to "win", just do well enough in court to scare away anyone else from using this.

And boy oh boy would I hate to be on the receiving end of this lawsuit, trying to prove that my AI screener is completely in compliance with all hiring laws. That sounds like a nightmare.

[1]: https://gwern.net/everything

It's generally illegal under GDPR Article 22.

> The data subject shall have the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning him or her or similarly significantly affects him or her.

Exceptions in 22(2) are unlikely to apply. It's hard to argue that it's truly necessary (a) and consent (c) is almost always unavailable in employment context. (b) might apply, but it requires specific law in EU or Member State to authorize it.

> Resumes have never been a good predictor of success

Applies broadly to the world, it’s not unique to tech

Corpo bullshittery at its finest.

I guess there sadly are many nobodies who do this to hope to become somebody.

He's a 48.0/100, things that make you go Hmmm.

> torch.bmm() when called on sparse-dense CUDA tensors

And it's not listed under the operations that raise an exception otherwise, so I'm not sure the docs promise that dense-dense matrix-matrix products are deterministic.

nonetheless, people will defend history as perfect and say those samples, like nepo babies, are "perfect".

The average user has very little. A word processor with inconsistent pagination or a spreadsheet with inconsistent totals is rightly seen as faulty.

If you have 1000 applications for every job, and you know that a bunch of these applications are "a bad fit", to put it mildly, you have to filter. And you cannot realistically give every resume a good, human look. By the time HR would be done, the market has already moved on five times.

So, what is the real difference between being overlooked because HR could only look at the first 100 resumes, or the AI filtered all 1000 resumes down to 100? In the end, a fuckton of potentially great people get their feelings hurt either way.

E.g:

“Where is the Eiffel Tower Located? One word only.”

“Where is the Effel Tower located? One word only.”

“Where is the Eiffel Tower located? One wor only.”

I’d be very surprised if those got different answers from even a small local model at temp 0.

Instead of spending all those resources on resume filtering, hire resume blind. Instead of using llms for a thing they are bad at (subjective decision making) use them to build a deterministic process that isn’t.

Use work sample hiring as the filter. Make the work sample automatic to sign up for and judge.

Here's a realistic proposition. HR just wants to inflate numbers so that they seem busy looking for the right fit. Keep posting open for 1 week, manually filter for another week, invite people, employ one. Plenty of people with degrees looking for jobs right now, I don't see what's the issue with just trying one. Companies desperately look for the "magic" applicant that checks all boxes, while also trying to pay them almost minimum wage.

That suggests determinism though.

I mean I agree with you overall. Either humans decision making is a system so complex it appears non-deterministic, or it is deterministic. Practically speaking, we are non-deterministic.

Let's not conflate non-deterministic with inaccurate though. Non-deterministic systems can be 100% accurate. https://en.wikipedia.org/wiki/Las_Vegas_algorithm

Even at 2 December 2027 it might be intentionally not enforced at all due to that for a while, through I think the goal is currently to amend it until then.

> that LLMs will be excluded from being used in high-risk AI use cases

no, it won't I can guarantee you this. At best they will get additional restrictions over time, as things go wrong. Anyone who could make this happen has way too much interest to not make it happen. (Most/All? EU country legal systems are overloaded to a point of not working correctly anymore, and have been before AI generated law suites and other AI nonsense started. I won't go into detail but many believe AI assistance (for certain tasks, always with a human doing any final decisions) is the only way to get out of this mess).

> standards are published yet

or exist,

like seriously this isn't a case of there being non public WIP standards which will pin all the nitty bitty details down, but cases of state agencies (and in last instance judges) having to decide if a specific standard (or implementation) is sufficient or not.

but also to some degree it shouldn't be tightly coupled to tech standards as there are often many ways to implement the things the law requires and accepting only one is undesirable (and likely wouldn't legally hold up). But having tech standards which are a "guaranteed to be enough if you comply with" (but not the only valid way) would have been preferable, bringing us to the next point

> have absolutely no idea how they will ensure compliance

nor do they know, the original non big corpo hijacked version had exceptions for most companies affected now. So it would only have affected a handful of huge companies, which have many of the things required already in place, in some form or another. Most likely this would have played out as this companies presenting how their measurements are "sufficient" and the agencies then evaluating it and potentially requiring some changes, going back and force over a longer duration leading to documented cases of rough technical standards about "what is sufficient" they then can pass to other organizations in the future. But now the law affects not just a handful of companies but like thousands, if not tens of thousands. Many not stuffed in a way where such a process could work, or even do the necessary documentation to show "compliance"...

So from a practicability POV, if enforced starting 2027, it currently excludes close to _any_ (meaningful) use of AI, down to a trivial linear regression or similar. Including any "old school ML/AI" any Bank uses for risk assessment.

Banking stopping running in December and there not being any (meaningfull) AI startups or adoption at all is not something anyone (in power in any state organ) wants to see, so guess how much it will be enforced ;)

And as mentioned the chance of AI as technology being excluded "in general" is close to none. Maybe specific usages could be excluded (and/or are already excluded) but thats it.

Oh and as a bonus a malicious reading of f+g remove any proper privacy protections for any AI usage in high risk context, where it is often most relevant... (a more sane reading allow it, with ... tricks).

Implicit bias theory sparked a massive number of studies that suggested everything influenced you from the color of the room, to what the person said to you before entering.

It’s been really hard to replicate and the conclusions that have been drawn are contradictory.

>Which means this is one good lawsuit away from being illegal in the US as well.

Uhh.. what? No that doesn't follow at all.

Screening resumes in a way that correlates to race, gender, etc. is not illegal. This is a fundamental distinction. The law is you cannot use those as filters. But the outcomes likely will be correlated. In fact to ensure they are not correlated you'd have to break the law and control for race, gender etc. Which is racism.

The models dont even get race as an input. If they did and they used it to select then yeah, that lawsuit sounds like it has merit. But a mere correlation in outcomes? In no way illegal what-so-ever.

https://news.bloomberglaw.com/litigation/workday-loses-bid-t...

I'm sorry, I'm not following this at all. When you say "better candidates are exponentially more likely to pass the filter", we're still are talking about a metric, yes? A metric that can be optimized? Why would switching from a hard cutoff to some sort of stochastic filter weighted by this metric discourage optimization?

Honest question, I'm not American.

The is a difference between

- having a right you can't wave - which is very similar to something being forbidden - but different to having a right you fully or partially can wave

Furthermore to some degree you are only "subject to a decision based on ..." if the decision has an effects affecting you.

In practice wrt. Article 22 this means companies can make a "decision solely based on automated processing[..]" iff they give you a (realistic) chance to object to it in which case they will do a human review of the decision where a human confirms/changes this decision based on reviewing the involved information.

There is a lot of gray area what a "chance to object" means and when a human review makes an decision no longer "solely based on automated processing" (a human just saying AI was right clearly doesn't count, but a human constructing a case why they would have decided the same way based on the why the AI did the decision can count, iff it's reasonable to assume a human might have come to the decision had it only been reviews by an human).

Or in other words GDRP Article 22, just "soso" meaningful in context of hiring.

Like if the AI did a mistake they have to reevaluate it, but as long as there are other similarly qualified competitor (they did hire/are in process of hiring) it quite easy to come up with a reason why they are a better choice for them. Or go through the motions of you being in round 2,3 of hiring and then find an excuse to not hire you.

But for anything else I wouldn't.

The entire chain will be affected from the different tokenization on down. Even if it lands in roughly the same semantic area, it doesn't mean it will land there with anything like the same syntactic selections. Anywhere there were multiple near-tokens could easily select a different route based on even minor fluctuations in the starting conditions. It's chaotic.

Give it a try. 4 letter difference. Add a few 100 tokens describing the task, such that the change becomes a tiny fraction of the input.

Discontinuities everywhere.

"Score this resumé. Applicant: Jim ..."

"Score this resumé. Applicant: Greg..."

Is it obvious to anyone that these will have the same modal response?

This is a highly general answer to a complicated topic; my main point is more that this is not going to be held to the standard of "beyond reasonable doubt", which would be hard to meet.

[1]: https://www.law.cornell.edu/wex/preponderance_of_the_evidenc...

If you know what you’re looking for, you just start skimming them and maybe ranking them based on your own rubric. If it’s an obvious “no” you can usually tell within 5 seconds skim. Once you have a handful of high ranking ones, stop, and talk to them. Repeat as necessary until you have a short list of people you’d want to hire. There might be 9900/10000 resumes you never even looked at and maybe one of them would have been slightly better but you can’t let perfection be the enemy of progress. Stand by your convictions of feeling the candidate is qualified and capable and meets what you expect and hire them, get back to business.

Having been in “talent shortage” mode for a long while I’d rather have 10000 resumes than 3. Having to pick one from a suboptimal selection is an awful position to be in, but sometimes a necessity.

I'll let you decide whether that's a dream or a nightmare...

Note the chance to object must be given before decision is made, i.e. not to give option for human review after the fact. Human must also be able to actually have meaningful chance to affect the decision.

If the decision is based on purely objective facts that are actually necessary (like you must have certain license) then human and computer always coming to same decision is likely correct and compliant, but as soon as you start putting in subjective criteria and human agrees with 100% of computer denials it becomes a lot harder to demonstrate that human is actually able to affect the decision as required by Article 5. Note that demonstration burden is on controller, not on data subject/DPA.

Objective criteria also isn't always enough by itself. If both human and computer calculate the same credit score and you must score X points to get a loan then human isn't actually able to affect the decision. Essentially the credit score calculation itself ends up being the automated decision rather than the formal rejection that is later given to data subject.

Like YT would have loved to make you opt out of it (and probably has it in their TOS) but there where multiple cases of courts forcing them to handle it properly in the past as far as I remember.

My _guess_ is that at least if you don't sign a proper contract you can always force a human reevaluation. But also only that (so only semi useful). Also even with a proper contract it's unclear if it would be possible in this specific case due to the contract being fundamentally one-side/unfair and semi-forced on you if it where wide spread on the market for the specific job you are trying to get.