Telling GPT-4 you're scared or under pressure improves performance

In the case of flight we're interested in a function: transportation. The plane performs that function so we call it 'flight'.

Were we interested in navigating the world as a flying animal does: in flocks, social navigation, hunting, etc. then indeed, planes would not count. Planes, in this sense, do not fly. Planes, in this sense, are a trick.

We already have 'functional intelligence', we've had it for thousands of years and perfected it in the 20th C with the electronic computer. Any system of automation we build is 'functionally intelligent', including the plane.

The problem is were not interested in this form. When Commander Data was written, with "I, Robot" and the like, the authors were writing People. They were writing animals. They imagined flying birds, just made of metal.

This is a (nomenological) impossibility, just as impossible as making an aeroplane to fly like a bird.

The kind of intelligence which matters to us is animal intelligence. The kind which matters to engineers, of course, is purely functional: more trinkets to sell.

But you cannot glue these together and call them a person, nor pronounce that what matters to you is the same as what matters to everyone else. This is a delusion, or a lie, and largely a mixture of both. A series of lies told to the public to maintain a popular delusion, one which is profoundly dangerous.

A creature in the world with us, talking to us and meaning what it says, judging situations we are in, advising us based on our needs and its understanding of them -- etc. is a creature whose mode of operation and mode of life is alike our own.

Lying that LLMs are this, saying that 'planes flock together in the sky', is dangerous. Users of these systems adopt a schizophrenic disposition to them, and rely on them -- this reliance, based on a lie, is dangerous to them. These systems have no such capacities. They generate text according to what is, on average, best given a historical corpus.

They do not imagine, reflect, emote. They do not move, sense, or coordinate. They have no intention to speak, and cannot mean what they say. They are not here with us. 'They' are not a 'they' at all -- rather, cleverly constructed tea-leaves based on a shredded recording of everything ever written.

In the matters of intelligence, we want an animal -- we do not want a calculator. This is a solved problem. And it is a dangerous thing to tell people their calculator's advice has considered their interests -- what nonesense.

I suspect we are getting into an issue of degrees, potentially due to differences in how you and I have been applying LLMs.

For example you said that in the average case they actually represent concepts by the complex interplay of adjacent concepts - I would agree. ChatGPT can pass the bar, it can pass medical exams etc. I would also point out that the work in that sentence is being done by the term “average case”.

Let’s assume our experiences diverge at this point. At the start of the year, I started tinkering, then actively trying to push LLMs to failure, in order to understand the limits of what could be achieved.

After creating several tools/experiments you end up having to deal with Hallucinations, and this is where my stance likely diverged from yours.

Two different studies showed generated content was only ~50% and ~40% supported by provided citations.

One out of 4 of my summarization tests was spectacularly fabricated.

I had bad performance on even classification tasks - and OpenAI engineers described this same failure at a conference. I am a recovering non-coder, so you dont have to take my word for it.

At work, I need processes that are more than ~97.x% accurate, otherwise they are poor replacements for the human in the loop ones already in place.

Average case performance suggested the ability to actively plan, to actively assess situations. However hallucinations overrode those capabilities. LLMs will actively imagine functions, teams, or processes that dont exist.

Eventually, it became clear that LLM hallucination is far too anthropomoprhized a word. LLMs are always “hallucinating” - it’s only humans who have an issue with the output.

If I have understand you correctly, semantic inaccuracy to you is simply a failure of not having enough related concepts.

I wish I could remember the exact examples that made me realize there is no world view at play at all, I could simply share those.

Instead, can you describe how you are getting acceptable performance from your LLMs? Maybe experience and use cases will be enough to bridge the gap.

No, because I have reasons to type and it is those reasons which express what I am doing.

An LLM has no reason to be doing anything; it is not responsive to reasons.

If I say, "i like what you're wearing" i may: be flitring, expressing my taste, being enouraging, etc -- perhaps all at once.

It is precisely all these reasons for action which LLMs lack. They generate text on the occasion of a prompt, not for any reason (in this sense) at all. So literally: they do not act.

An LLM is more like a river than a person. The flow of electrons which brings about a response to a prompt is a (very narrowly) deterministic function of a historical corpus of text.

Whereas a person is a narrowly non-deterministic, or very broadly deterministic, function of their experiences and capacities. People grow in their environments, and in growing, acquire novel dispositions which give them reasons for acting.

The word "just" here is very important. They are, very much, just generating text.

There really isnt any significant achivement here at all. Big tech companies stole decade's worth of our electronic data --- comments, books, forums we created to share with each other -- and ran it through many-mil-$ hardware costing many-mil-$ electricity. They ran it through a fundamentally simple algorithm.

All the achivement here is ours as a species communicating digitally and recording our lives. I regard OpenAI, et al. as profoundly parasitical on this. Replaying ourselves back to us, and claiming "ChatGPT" as an author.

This scam-framing hoodwinks investors, and the public, into ever higher valuations based on ever more ridiculous hagiography. There is a tool here, and it's value comes from us

My reading of the papers is, given enough scale, modality, and memory; there are chances (perhaps newer and different) models will be able to "generalize" our world. Also: https://archive.is/3yyZZ / https://twitter.com/QuanquanGu/status/1721394508146057597 | And: https://archive.is/bW2tS / https://twitter.com/mansiege/status/1680985267262619648

Spell check is crazy inconsistent and poorly implemented. I'm constantly running across words the browser says is misspelled but isn't. I know you can replace the built in spell check with a better dictionary, but trying to keep that in sync across all my devices just keeps me going back to Google as spellcheck.