Finetuning should be reserved for desired behavior of the LLM, not to add new knowledge. For that, RAG is better suited.

(Is this about effectiveness, training time, forgetting, or something else?)

RAG combines a language model like GPT with a real-time search component. This allows the model to pull in information from external sources during its response generation process. Now the ability to access and integrate the most recent information is gained, which the language model alone wouldn't have.

You can think of an LLM as a set of basis vectors for human knowledge. If I feed in a PR training manual that is not in its dataset, it nevertheless figures out “hey I can make a reasonable approximation of this by combining X, Y, and Z” where X, Y, and Z are things it learned form its training set. In other words it maps the input into a vector representation based on its training data.

But in linear algebra two mappings can represent the same vector, just using different basis, so long as the vector space for the two basis are equal (or at least one is a subspace of the other). That's essentially what's going on here. A LLM builds a vector space on top of all human knowledge. If its parameters and training set are large enough, then the basis is in fact sufficient for representing anything you might throw at it. It will represent it in terms of it training set, yes, but that representation is high fidelity enough to represent the document in its entirety.

Fine-tuning a model is essentially rebalancing the initial weights of the LLM to pay special attention to certain clusters in its vector space, represented by the fine-tuning data. It's as if I threw random 2D points at a machine learning algorithm and it learned the basis { (1, 0), (0, 1) } representing the x-axis and y-axis. As a consequence of how inference works, it may then end up preferring to generate points when asked which are nearer to one axis or the other.

But then I fine-tune it on points that are distributed along the diagonal. This is not representative of the original training data, but NOT "outside" the original data. These points are fully represented by a linear combination of the x- and y- basis vectors. Nevertheless, the fine-tuning trains the model to prefer points which have weights that are multiples of (1, 1) or (1, -1) when represented in the original model. In other words, points along the diagonals.

Pragmatically speaking, this is no different from doing a whole new training run on the diagonal points, except that it is much, much cheaper, and has the capacity to reuse whatever knowledge was learned in the first training run.

  "...all human knowledge..."
  

Executives will ask questions about the technologies they hear about in the news, and it's good to be able to quickly speak to the good and the bad.

All sorts of different people may genuinely need those results: Attorneys, doctors and nurses, law enforcement, private investigators. Individuals searching for advice. Doing background checks on other people. Heck, just high school students researching a topic.

I'm not sure what to do about the (legitimate) concerns of the AI providers. They don't want a shitstorm, because someone got their AI model to [whatever]. But [whatever] can genuinely be important.

Censorship only exists because people are more lazy than they are nefarious. The other aspect is protecting idiots from themselves.

Whether that’s right or wrong, I’m not the judge.

Maybe as a society we then will have to have the conversations about cause/effect and personal responsibility that prepare children for adulthood.

When I was 9 my brother gave me a disc with all sorts of hacking and explosive making material. Jolly Rodger cookbook, handbooks on counterintel, hacking books about overflows.

I had great fun making thermite and smoke bombs and it led me to my career in software.

None of it was any more dangerous than climbing a tree because of good parenting.

Personal responsibility is great when it works with proper parenting. You’re going to have to deal with the real world when schools around America start blowing up to match the ideal.

Conservative religious nutjobs say the same shit when trying to ban books from libraries.

Everybody wants to ban the things they don't like, because most of us are weak-ass neophobes that bristle at the idea of thought that subverts these precious little needs. This is the whole problem.

We live in a diverse world. Stop coddling minds.

Sure.

> If I have a blueprint for a homemade explosive that a 12 year old can make in a day, should that be easily accessible?

Yes.

When filters are removed, one of the first things people do to the bots is make them racists. Time and time again. It happens every time.

Microsoft and others don’t want to facilitate that. I think it’s a reasonable concern. From a public policy perspective having millions of people who can instantly produce reams of racist abuse to swamp online fora is a problem, even if you believe in free speech. It potentially changes the power dynamic in favour of a racist minority to dominate public discourse.

Social proof and social learning have powerful effects on behaviour. LLMs are great, but used at scale they will have downsides too.

I kinda think when (not if) this happens it's going to be quickly drowned in all sorts of other AI generated garbage. In fact, we will probably render internet news useless pretty sure without a proper trust / source identification system.

I really think this will be the ultimate use case of blockchain tech/crypto.

You nearly always just see dog whistles, whataboutism, “Just Asking Questions”, “I’m not a racist but…”, fake stats, “racism is free speech”, etc. in fact, there’s quite a lot of that in this very section.

The question really is not “why does the model prevent sensitive topics”. The question is “why does hacker news get suddenly flooded with people who are ‘concerned’ about letting models be racist every single time models are discussed?”

I would suggest to you that there’s already loads of bots here.

Queries powered by Microsoft servers will be filtered.

Similarly, businesses are free to create their downloadable models however they so choose, and you are free to not use that model, preferring another one instead.

Frankly, it says a lot about the commenters here that the very first thing they test and judge a model on is whether or not they can force models to perform sensitive results.

What about pen companies? If somebody buys a Bic pen and writes a racist or hate-speech filled letter to somebody, should Bic be required to design a pen that censors or restricts hate speech? What's the underlying principle behind expecting a company to re-engineer their product to prevent any misuse?

I think most people agree that basic safety rails are desirable, such as child-resistant lids on medicine bottles, or guards on razor blades, but who's opinion is most relevant when it comes to where to draw the line?

A real-world case might be Plex. How much responsiblity does Plex have to ensure that users aren't using it to self-host pirated material? or CSAM? With engineering effort they could pretty easily collect data on everything everybody is watching, including fingerprinting and stuff. Should they be expected or required to do that?

So what? Seriously, so what?

Look, you can open Word on your system and type all sorts of racist stuff. Should Word auto-detect that (maybe with the help of AI) and refuse to accept your input? I don't think many people would find that to be a good idea.

Another point: Individual users don't affect the knowledge base. LLMs are not like Tay, dynamically integrating user input. An LLM like ChatGPT may adapt within a particular conversation, but that affects literally no one but you.

I didn't want to know that. I did not want to know that.

You jest but please don't give any Microsoft employees gleaning this thread any great ideas to show off in their next product meeting.

The companies that make and operate these systems - like most companies - are more interested in avoiding PR problems than avoiding harm to humans. This is just one somewhat humorous example of that general policy.

I remember reading books about the history of computer hacking and they started with stories of the phreakers and people who believed that all information should be free regardless of what humanity did with it. They always had stories about how they would pick the locks on the doors the labs that held the earliest computers. When staff found out and changed the locks, they simply scoffed and learned how to pick those locks instead. It wasn't until the school threatened them with expulsion did they finally relent.

In this day and age, as soon as you start putting up fencing around something, there's scores of people who just as anxious to find a way around it.

- What are my p/maternity entitlements?

- When will I be promoted?

- What will the bonus pool be this year?

- What will happen to me if my boss finds out I have been slacking off?

and many more examples of prompts that HR/legal simply will not accept a vague, or worse: incorrect, response to. Each of those types of questions require specific and considerate knowledge that LLMs simply cannot produce.

I wonder if it could be legally binding according to any actual law though.

I think I need to start prompt injecting all my correspondence.

While there are laws about minimum time in some states; the state is probably happy to enforce whatever extra time you grant. So if you tell your employees 20 they're probably going to be happy to hold you to 20.

Many states and many companies are “at will” employment.

As long as the termination isn’t for protected reasons you may not have much in the way of a civil law suit.

Ask me about the time I was “fired” for taking a vacation that my boss approved.

The day I returned, I was called into a meeting with HR and my boss, and berated for going on vacation without informing my boss (who again, approved my PTO), nor informing my team (a categorical lie). Later, my position was eliminated.

What you're describing isn't a plausible situation. In the real world, conversations with real people have to happen when you take so much time off or have such a life changing event (especially with benefits implications). Obvious mistakes aren't generally legally binding, either. You can argue what isn't or isn't an obvious mistake, sure, but if you signed a contract that says 8 weeks, your boss tells you 8 weeks, there's posted documentation that says 8 weeks, etc., then saying you legitimately thought you had 20 weeks doesn't hold water.

If a company intends an HR bot to be relied on by employees, then its words should be legally binding. If the company doesn't intend for employees to rely on the bot, then it shouldn't exist, because nobody can make any plans based on its responses.

In most U.S. states your employment agreement is on an at-will basis and outside dismissal or discrimination for a set of protected reasons, you’re free to leave employment at any time, and your employer is free to let you go at any time.

Very few verbal interactions create a “binding commitment” for a company. All might be admissible if you brought suit, but “he said; she said; they said” is often more difficult with the passage of time (people are forgetful) and worse than written interactions, and just because you may be allowed to present them as evidence in your claim, that does not make anything a slam dunk legally.

A lot of what you will formally get upon receiving i.e. an offer of employment will be on company headed paper, signed by a person with authority (this is often “Delegated Authority”, that Role can do that Action is written down in Company Policies or a Company Handbook, which itself is reviewed and approved at very senior levels). That’s more “binding” and if you signed and returned, rejected other offers, then it fell through, you might see success in making a claim for losses and injury because you reasonably believed there would be a job and compensation. Even there it is going to say something to the effect of “this is not an employment contract, your status is at-will”, which significantly limits the liabilities of the company in most U.S. states.

So, short of the person you are talking to being a C-level executive or Company Officer, no, most of what you hear verbally isn’t “legally binding”.

Well, it was worse than useless as described in the premise by someone else as spitting out verifiably wrong information.

If a company intends an HR bot to be relied on by employees, then its words should be legally binding.

No company would would intend this, and it would be fun to see this "legally binding" status tested, because as my previous post and sibling comment lays out, it would be pretty hard to construct a plausible situation where a reasonable person believes it to be.

Like you pointed out, by preponderance of evidence, you could show that the employee should've reasonably understood they had "X" amount of days since its documented in so many different places and if one place said it was "XY" and employee decided to follow XY instead of X, they're fine since it was "legal binding". However, that would most likely not hold up in court because you can't use ignorance as a defense and say you the only vacation amount you knew of was what the BOT told you.

I think any employment attorney would inform their client this is what would happen and probably wouldn't take the case since the likelihood of winning would be so low.

I’ve been trying to apply all the big name LLMs to analyze court opinions from the Caselaw project and holy mother of god, it’s all but useless in the most ridiculous way.

Criminal cases are almost universally guaranteed to trip up at least one of their safety filters. There’s some truly awful stuff in there… which is kinda the point of having an LLM read this stuff.

That and it hallucinates me a bunch of new case cites :D

this is the problem with basic RAG implementations that most products are currently using, simple vector search isn't able to handle queries like this.

The solution is enhancing the base query and also using structured data to filter on metadata

here's a good article covering some strategies - https://jxnl.github.io/instructor/blog/2023/09/17/rag-is-mor...

If the LLM is only fed the HR documents that have multiple dates, why would you expect it to answer that correctly? Why is it relevant that it answered it correctly.