But everyone is expecting them to release gpt5 later this year, and it is a bit scary to think what will it be able to do.

The evidence for that is the change in the tokenizer. The only way to implement that is to re-train the entire base model from scratch. This implies that GPT 4o is not a fine-tuning of GPT 4. It's a new model, with a new tokenizer, new input and output token types, etc...

They could have called it GPT-5 and everyone would have believed them.

The expectations for gpt5 are sky high. I think we will see a similar jump as 3.5 -> 4.

I assume GPT-5 has to be a heavier, more expensive and slower model initially.

GPT-4o is like an optimisation of GPT-4.

1) It's natively multi-modal in a way I don't think gpt4 was.

2) It's at least twice as efficient in terms of compute. Maybe 3 times more efficient, considering the increase in performance.

Combined, those point towards some major breakthroughs having gone into the model. If the quality of the output hasn't gone up THAT much, it's probably because the technological innovations mostly were leveraged (for this version) to reduce costs rather than capabilities.

My guess is that we should expect them to leverage the 2x-3x boost in efficiency in a model that is at least as large as GTP4 relatively soon, probably this year unless OpenAI has safety concerns or something, and keeps it internal-only.

Almost everyone here has used at least one LLM for fun or for work.

Many people here have paid for it.

They also said it about all the over-hyped tech that did not change the world. This mentality of “detractors prove something is good” is survivorship bias.

Note I’m not saying you’ll categorically be proven wrong, just that your argument isn’t particularly strong or valid.

LLMs are no different. Like DL modules that are very good at outputting images that mimic biological signatures, LLMs are very good at outputting texts that eerily mimic human language.

However — and this is a point which programmers are woefully and comically ignorant — human language and reason are two separate things. Tech bros wholly confuse the two however, and thus make outlandish claims we have achieved or are on the brink of achieving — actual AI systems.

In other words, while LLMs and DL in general can perform specific tasks well, they do not represent a breakthrough in artificial intelligence, and thus will have a much narrower application space than actual AI.

LLMs statistically imitates texts of real world. To achieve certain threshold of accuracy, it turns out they need to imitate the underlying Turing machine/program/logic that runs in our brains to understand/react properly to texts by ourselves. That is no longer in the realm of the old school data-as-data statistics I would say.

You're talking more about AGI.

We need "that's not AI" discussions like we need more "serverless? It's still on some server!!" discussions.

It's about meaning of intelligence. These people don't have problems claiming that ants or dolphins are intelligent, but suddenly for machines to be classified as artificial intelligence they must be exactly on the same level as humans.

Intelligence is just about the ability to solve problems. There's no implication that in order for something to be intelligent it has to perform on at least the same level as top people in that field in the World.

It just has to be beyond a simple algorithm and be able to solve some sort of problem. You have AIs in video games that are just bare logic spaghetti computation with no neural networks.

In other words, you’ve lost the argument.

Just stop trying to redefine AI as a term, you'll lose against the old hands and you'll lose against the marketing dept and you'll lose against the tech bros and nobody who you actually need to explain it to will care. Use AGI or some other common term for what you're clearly talking about.

... in the human brain which has evolved "cores" to handle each task optimally.

It's like the Turing Test. If it looks like it's reasoning, does it matter that it's doing it like a human brain or not?

The method is irrelevant. The output is what matters.

This is like a bunch of intelligent robots arguing that "mere meat" cannot possibly be intelligent!

https://www.mit.edu/people/dpolicar/writing/prose/text/think...

No one is talking about it being AGI. Everyone is talking about just AI specifically. I think your problem is thinking that AI = AGI.

For example AI in video games is very specific and narrow to its domain.

What a bizarre claim. If LLMs are not actually outputting language, why can I read what they output then? Why can I converse with it?

It's one thing to claim LLMs aren't reasoning, which is what you later do, but you're disconnected from reality if you think they aren't actually outputting language.

It will give you detailed examples that are much easier to follow than vague, error-prone spec docs. That's scratching the surface. Other people are far more creative than me and have used chat-gpt for mind-blowing stuff already. Whatever its doing passes for 'reasoning' and 'intelligence' in my book. To me it doesn't matter whether its the same kind of intelligence as a human or if there's any amount of awareness as those are both philosophical questions of no consequence to my work.

For what these pieces of tech can do I feel that they're drastically under-utilized.

After an OpenAI launch, I think it's important to take one's feelings about the future impact of the technology with a HUGE grain of salt. OpenAI are masters of hype. They have been generating hype for years now, yet the real-world impacts remain modest so far.

Do you remember when they teased GPT-2 as "too dangerous" for public access? I do. Yet we now have Llama 3 in the wild, which even at the smaller 8B size is about as powerful as the [edit: 6/13/23] GPT-4 release.

As someone pointed out elsewhere in the comments, a logistic curve looks exponential in the beginning, before it approaches saturation. Yet, logistic curves are more common, especially in ML. I think it's interesting that GPT-4o doesn't show much of an improvement in "reasoning" strength.

It's glib to dismiss safety concerns because we haven't all turned into paperclips yet. LLMs and image gen models are having real effects now.

We're already at a point where AI can generate text and images that will fool a lot of people a lot of the time. For every college-educated young person smugly pointing out that they aren't fooled by an image with six-fingered hands, there are far more people who had marginal media literacy to begin with and are now almost defenceless against a tidal wave of hyper-scaleable deception.

We're already at a point where we're counselling elders to ignore late-night messages from people claiming to be a relative in need of an urgent wire transfer. What defences do we have when an LLM will be able to have a completely fluent, natural-sounding conversation in someone else's voice? I'm not confident that I'd be able to distinguish GPT-4o from a human speaker in the best of circumstances and I'm almost certain that I could be fooled if I'm hurried, distracted, sleep deprived or otherwise impaired.

Regardless of any future impacts on the labour market or any hypothesised X-risks, I think we should be very worried about the immediate risks to trust and social cohesion. An awful lot of people are turning into paranoid weirdos at the moment and I don't particularly blame them, but I can see things getting seriously ugly if we can't abate that trend.

I guess we need to have an AI secretary to take in all phonecalls from now on (spam folder will become a lot more interesting with celebrity phone calls, your dead relative phoning you etc)

It's hard to know what to believe if you get a phone call with the voice of your child or colleague, and your "child"/"colleague" replies within milliseconds in a convincing way.

I think unfortunately it will massively lower the trust of actual real videos and images, because someone can dismiss them with little thought.

As someone that grew up with late-90’s internet culture and has seen all the pros and cons and changes over the decades, I find myself using the internet less and less for dialogue with people. And I’m spending more time in nature and saying hi to strangers in reality.

I’m still worries about the impact this will have on a lot of people’s ability to reason however. “Just” Tik Tok and apps like it has already had devastating results on certain demographics.

Brave New World indeed.

Chances are they've already shoved some app on their phone that's voice to txting everything they say and sending off somewhere (well lower chance if they have an iphone).

Modern life is data/information security and humans are insanely bad at it.

I don't think it's a problem to find a word or a sentence or a story - whatever - that's commonly used by everyone on daily basis but in different context. That's not a problem by itself :) try it

For the idiots, it is still possible to find a word. They may be idiots, but still, they work and live on their own. They coming along in life. So, it's up to the smarter one to find a no-brainer solution.

I am confident and believe nothing and no one is stupid enough not to be able to adapt to something. Even if it's me, who'll need to adapt to members with less brain.

This is my biggest gripe against the telecom industry. Calls pretending to be from someone else.

For every single call, someone somewhere must know at least the next link in the chain to connect a call. Keep following the chain until you find someone who either through malice or by looking the other way allows someone to spoof someone else's number AND remove their ability to send the current link in the chain (or anyone) messages. (Ideally also send them to prison if they are in the same country.) It shouldn't be that hard, right?

The technology to build trustable caller ID has existed for a long time, the problem is no one wants to be the one forcing telcos all over the world to upgrade their sometimes many decades old systems.

And that's the case even if you've never ever posted anything on your social media - it could be family&friends, or employer, or if you're ever been in a public-facing job position that has ever done any community outreach, or ever done a public performance with your music or another hobby, or if you've ever walked past a news crew asking questions to bystanders of some event, or if you've ever participated in some contests or competitions or sports leagues, etc, all of that is generally findable in various archives.

I'm sure AI-based ageing can do a good enough job to convince many people that a fake image of someone they haven't seen for years is an older version of the person they remember; but how often would it succeed in ageing an old photo in such a way that it looks like a person I have seen recently and therefore have knowledge rather than guesses about exactly what the years have changed about them?

(Not a rhetorical question to disagree with you, I genuinely have no idea if ageing is predictable enough for a high % result or if it would only fool people with poor visual memory and/or who haven't seen the person in over a decade.)

I feel like even ignoring the big unknowns (at what age, if any, will a person start going bald, or choose to grow a beard or to die their hair, or get a scar on their face, etc.) there must be a lot of more subtle but still important aspects from skin tone to makeup style to hair to...

I've looked up photos of some school classmates that I haven't seen since we were teens (a couple of decades ago), and while nearly all of them I think "ah yes I can still recognise them", I don't feel I would have accurately guessed how they would look now from my memories of how they used to look. Even looking at old photos of family members I see regularly still to this day, even for example comparing old photos of me and old photos of my siblings, it's surprising how hard it would be for a human to predict the exact course of ageing - and my instinct is that this is more down to randomness that can't be predicted than down to precise logic that an AI could learn to predict rather than guess at. But I could be wrong.

Do you shop in large corporate stores and use credit cards? Do you go out in public in transportation registered to you?

If yes, then images and habits of yours are being stored in databases and sold to data brokers.

And you're not even including every single one of your family members that use internet connected devices/apps that are sucking up all the data they can.

But don't disagree with you - in a different comment that was about privacy, I (despite living under GDPR) suggested that for offline verification with known people it's better to choose secrets that definitely haven't been shared online/anywhere rather than just choosing random true facts and assuming they couldn't have been found out by hackers: https://news.ycombinator.com/item?id=40353820

Why not an AI assistant in the browser to fend all the adversarial manipulation and spam AIs on the web? Going online without your AI assistant would be like venturing without a mask during COVID

I foresee a cat-and-mouse game, AIs for manipulation vs AIs for protection one upping each other. It will be like immune system vs viruses.

We went from living in villages where everyone knew each other to living in big cities where almost everyone is a stranger.

We went from photos being relatively reliable evidence to digital photography where anyone can fake almost anything and even the line between faking and improving is blurred.

We went from mass distribution of media being a massive capital expenditure that only big publishers could afford to something that is free and anonymous for everyone.

We went from a tiny number of people in close proximity being able to initiate a conversation with us to being reachable for everyone who could dial a phone number or send an email message.

Each of these transitions caused big problems. None of these problems have ever been completely solved. But each time we found mitigations that limit the impact of any misuse.

I see the current AI wave as yet another step away from trusting superficial appearances to a world that requires more formal authentication protocols.

Passports were introduced long ago but never properly transitioned into the digital world. Using some unsigned PDF allegedly representing a utility bill as proof of address seems questionable as well. And the way in which social security numbers are used for authentication in the US is nothing short of bizarre.

So I think there are some very low hanging fruits in terms of authentication and digital signatures. We have all the tools to deal with the trust issues caused by generative AI. We just have to use them.

The best we can ever hope to do is find mitigations as and when problems arise.

Of course this is much more difficult when it comes to software, and very few serious people think the idea of a ever present government monitoring your software would be a better option then reckless AI development.

That's massive fast change, and we haven't culturally caught up to any of it yet.

This happened from the 15th century onward. By the 19th century more than half the UK population could read and write.

So the question, I think, is how do we reclaim trust in a world where every kind of content can be convincingly faked? And I think the answer is by rebuilding trust between users such that we actually have reason to simply trust the users we're interacting with aren't lying to us (and that also goes for building trust in the platforms we use). In my mind, that means a shift to small federated and P2P communication since both of these enable both the users and the operators to build the network around existing real-world relationships. A federation network can still grow large, but it can do so through those relationships rather than giving institutional bad actors as easy of an entrance as anyone else.

Isn't it rather brilliant that you can just ask questions of competent people in some subreddit without first becoming part of that particular social circle?

It could also reintroduce geographical exclusion based on the rather arbitrary birth lottery.

That's how I look at where we're going with AI. Plunge along into the new arms race first and build the capacity, then later figure out the treaties and safeguards which we hope will keep our society safe (and by that I don't mean a Skynet-like AI-powered destruction, but the upheaval of our society potentially as impactful as the industrial revolution.)

Humanity will get through it, I'm sure. But I'm not confident it will be without a lot of pain and suffering for a large percentage of people. We also managed to survive 2 world wars in the last century--but it cost the lives of 100 million people.

This a problem with all technology. The mitigations are like technical debt but with a difference. You can fix technical debt. Short of societal collapse mitigations persist, the impacts ratchet upward and disproportionately affect people at the margin.

There's an old (not quite joke) that if civilization fell, a large percentage of the population would die of the effects of tooth decay.

Discovering an asteroid full of gold, with as much gold as half the earth to put a modest number, would have huge impact to the labour market. Anything conductive like copper, silver, mining jobs would all go away. Also housing would be obsolete as we would all live in golden houses. A huge impact to the housing market, yet it doesn't seem such a bad thing to me.

>We're already at a point where we're counselling elders to ignore late-night messages from people claiming to be a relative in need of an urgent wire transfer.

Anyone can prove their identity, or identities, over the wire, wire-fully or wire-lessly, anything you like. When i did go to university, i was the only one attending the cryptography class, no one else showed up for a boring class like this. I wrote a story about the Electrona Corp in my blog.

What i say to people for at least 2 years now, is that "Remember when governments were not just some cryptographic algorithms?" Yeah, that's gonna change. Cryptography is here to stay, it is not as dead as people think and it's gonna make a huge blast.

The thing about cryptography and government is that it's easy to imagine for a great technology to be adapted on the governmental level because of its greatness. But it is another thing to actually implement it. We live in a bubble, where almost anyone knows about cryptographic hashes and RSA, but for most of the people it is not the case.

Another thing is that political actors are tending to try to concentrate power in their own hands. No way they will delegate a decision making to any form of algorithm — being cryptographic or not.

As soon as this becomes a problem, then it might start bottom-up, citizens to government officials, rather than top to bottom, from president to government departments. Then governments will be forced to formalize identity solutions based on cryptography. See also this case in Germany [2].

One example like that, is bankruptcy laws in China. China didn't have any law regarding to bankruptcy till 2007. For a communist country, or rather not totally capitalist country like China, bankruptcy is not an important subject. When some people stop being profitable, they will keep working because they like to work and they contribute to the great nation of China. That doesn't make any sense of course, so their government was forced to implement some bankruptcy laws.

[1]https://www.wsj.com/articles/fraudsters-use-ai-to-mimic-ceos... [2]https://news.ycombinator.com/item?id=39866056

All this would do is crash the gold price. Also note that all the gold at our disposal right now (worldwide) basically fits into a cube with 20m edges (its not as much as you might think).

Gold is not suitable to replace steel as building material (because it has much lower strength and hardness), nor copper/aluminium as conductor (it's a worse conductor than copper and much worse in conductivity/weigth than aluminium). The main technical application short term would be gold plated electrical contacts on every plug and little else...

.. And gold teeth and grillz.

I didn't know that copper is a better conductor than gold. Surprised by that.

[1] https://www.youtube.com/watch?v=E2Gd8CRG0cc

Maybe they didn't know, maybe none of their colleagues used it, their company didn't pay for it, or maybe all they need is an Excel update.

But I am confident that using Copilot would be faster than clicking through the sludge that are Microsoft Office help pages (third party or not.)

So I think it is correct to fear capabilities, even if the real world impace is still missing. When you invent an airplane, there won't be an airstrip to land on yet. Is it useless, won't it change anything?

Replacing all jobs except LLM developers? I’ll tell my hairdresser

But we do have a huge number of examples of jobs disappearing thanks to machines — even the term "computer" used to refer to a job.

More recently and specifically to LLMs, such losses were already being reported around this time last year: https://www.businessinsider.com/jobs-lost-in-may-because-of-...

And that motorized auto transport, it never went anywhere, it required roads. I mean, who would ever think we'd cover a huge portion of our land in these straight lines. Now, don't mind me, I'm going to go saddle up the horse and hope I don't catch dysentery on the way into town.

> it'd have to be impacting the real world

By writing business plans? Getting lawyers punished because they didn't realise that "passes bar exam" isn't the same as "can be relied on for citations"? By defrauding people with synthesised conversations using stolen voices? By automating and personalising propaganda?

Or does it only count when it's guiding a robot that's not merely a tech demo?

Apparently so: https://www.businessinsider.com/jobs-lost-in-may-because-of-...

Note that this article is about a year old now.

> Comparing it to the internet is insane, based off of its status as a highly advanced auto complete.

(1) I was quoting you.

(2) Don't you get some cognitive dissonance dismissing it in those terms, at this point?

"Fancy auto complete" was valid for half the models before InstructGPT, as that's all the early models were even trying to be… but now? The phrase doesn't fit so well when it's multimodal and can describe what it's seeing or hearing and create new images and respond with speech, all as a single unified model, any more than dismissing a bee brain as "just chemistry" or a human as "just an animal".

Maybe it's time to construct some (high[er] density) housing where people want to live? No? Okay, then maybe next decade ... but then let's construct transport for them so they can get to work, how about some new subway lines? Ah, okay, not that either.

Then I guess the only thing remains to construct is all the factories that will be built as companies decouple from China.

~ Paul Krugman, winner of the 2008 Nobel Memorial Prize in Economic Sciences

It's still early, and I don't see much in corporate communications, for instance, but it will be quite the change.

It's worse than I thought. They've already managed to mimick the median HN user perfectly!

We need one who's doing the dirty work of not discussing.

I fear that at some point the anonymity that made the internet great in the first place will be destroyed by this.

The dead internet theory started to look more real with time, AI spam is just scaling it up.

It has been so bad, I even considered injecting misspelling and incorrect grammar and bad punctuation into my prose to prove my words are mine.

i remember seeing the change when GPT-2 was announced

Set a memorable verification phrase with your friends and loved ones. That way if you call them out of the blue or from some strange number (and they actually pick up for some reason) and you tell them you need $300 to get you out of trouble they can ask you to say the phrase and they'll know it's you if you respond appropriately.

I've already done that and I'm far less worried about AI fooling me or my family in a scam than I am about corporations and governments using it without caring about the impact of the inevitable mistakes and hallucinations. AI is already being used by judges to decide how long people should go to jail. Parole boards are using it to decide who to keep locked up. Governments are using it to decide which people/buildings to bomb. Insurance companies are using to deny critical health coverage to people. Police are using it to decide who to target and even to write their reports for them.

More and more people are going to get badly screwed over, lose their freedom, or lose their lives because of AI. It'll save time/money for people with more money and power than you or I will ever have though, so there's no fighting it.

1. something you have

2. something you know

3. something you are

These three things are required for any authz.

At that point, how can you validate it, as there’s no shared secret? The only thing we had was validating childhood stories. After a preponderance of them, we accepted she was real (she refused to talk on the phone — apparently her voice was damaged).

We eventually met her in real life.

The point is, you can always use these three principles: asking relatives to validate the phone number — something you have — and then the stories — something shared — and finally meeting in real life — something you are.

I only say this sort of jokingly. Three out of four of my parents/in laws are questionably literate on the internet. It wouldn't take much of a "me bot" for them to start telling it the stories of our childhood and then that information is out there.

As technology advances those proportions will be boosted. Seems inevitable.

Alternatively while it may be difficult to trick you directly, phishing the passphrase from a more naive loved one or bored coworker and then parroting it back to you is also a possibility. 'etc.

Phone scams are no joke and this is getting past the point where regular people can be expected to easily filter them out.

I don't think anyone has a good answer to that question, which is the problem in a nutshell. Job one is to start investing seriously in finding possible answers.

>We need to roll back to "don't trust anything online, don't share your identity or payment information online"

That's easy to say, but it's a trillion-dollar decision. Alphabet and Meta are both worthless in that scenario, because ~all of their revenue comes from connecting unfamiliar sellers with buyers. Amazon is at existential risk. The collapse of Alibaba would have a devastating impact on Chinese exporters, with massive consequent geopolitical risks. Rolling back to the internet of old means rolling back on many years worth of productivity and GDP growth.

Well that's exactly the sort of service that will be extremely valuable in a post-trust internet. They can develop authentication solutions that cut down on fraud at the cost of anonymity.

This tech is dangerous, and I'm currently of the opinion that its uses for malicious purposes are far better and more significant than LLM's replacing anyone's jobs. The bullshit asymmetry principle is very incredibly significant for covert ops and asymmetric warfare, and generating convincing misinformation has become basically free overnight.

Even when it comes to people like our parents, there are things we would trust them to do, and things that we would not trust them to do. But what happens when you have zero trusted elements in a category?

At the end of the day, the digital world is the real world, not some seperate place 'outside the environment'. Trying to treat digital like it doesn't exist puts you in a dangerous place to be deceived. For example if you're looking for XYZ and you manage to leak this into the digital world, said digital world may manipulate your trusted friends via ads, articles, the social media posts they see on what they think about XYZ before you ask them.

The nature of this tech itself is probably what is getting most people - it looks, sounds and feels _human_ - it's very relatable and easy for a non-tech person to understand it and thus get creeped out. I'd argue there are _far_ more dangerous technologies out there, but no one notices and / or cares because they don't understand the tech in the first place!

The "yet" is carrying a lot of weight in that statement. It is now five years since the launch of GPT-2, three years since the launch of GPT-3 and less than 18 months since the launch of ChatGPT. I cannot think of any technology that has improved so much in such a short space of time.

We might hit an inflection point and see that rate of improvement stall, but we might not; we're not really sure where that point might lie, because there's likely to still be a reasonable amount of low-hanging fruit regarding algorithmic and hardware efficiency. If OpenAI and their peers can maintain a reasonable rate of improvement for just a few more years, then we're looking at a truly transformational technology, something like the internet that will have vast repercussions that we can't begin to predict.

The whole LLM thing might be a nothingburger, but how much are we willing to gamble on that outcome?

The world learnt to deal with Nigerian Prince emails and nobody is falling to those anymore. Nothing was changed - no new laws or regulations needed.

Phishing calls have been going on without an AI for decades.

You can be skeptical and call back. If you know your friends or family you should be able to find an alternative way to get in touch always without too much effort in the modern connected world.

Just recently a gang in Spain was arrested for "son in trouble" scam. No AI used. Most of the parents are not fooled in this.

https://www.bbc.com/news/world-europe-68931214

The AI might have some marginal impact, but it does not matter in the big picture of scams. While it is worrisome, it is not a true safety concern.

Probably why it's not released yet. It's unsafe for phishing.

has been for years mon ami. i remember when they started talking about GPT-2 here, and then seeing a sea-change in places like reddit and quora

quite visible on HN, esp. in certain threads like those involving brands that market heavily, or discussions of particular countries and politics.

- It helps them sleep at night if their creation doesn't put millions of people out of work.

- Fear of regulation

I second that. I remember when Google search first came out. Within a few days it completely changed my workflow, how I use the Internet, my reading habits. It easily 5 ~ 10x the value of Internet for me over a couple of weeks.

LLMs is doing nothing of the sort for me.

ChatGPT does this again for me. I am routinely getting zero useful results on the first page or two of Google searches, but AI is answering or giving me guidance quickly.

Maybe this would not seem such an improvement if Google's results were like they were 10 years ago and not barely usable blogspam

To me, this just sounds like Google Search has become shit, and since Google simply isn't going to give up the precious ad $$$ that the current format is generating, the next best thing is ChatGPT. But this is different from saying that ChatGPT is a similar step up like Search was.

For what it's worth, I agree with you that Google Search has become unusable. Google basically destroyed it's best product (for users), by turning it into an ad riddles shovelware cesspit.

That ChatGPT is similarly good like Google Search used to be, is a tragedy. Basically we had a conceptually simple product that functioned very well, and we are replacing it with a significantly more complex product.

The cool kids in those days used Metacrawler, which meta searched all the search engines.

And hey maybe when combined with GPT-4o AskJeeves will finally work as intended.

Metasearch only helps recall. It won't help precision, the metasearch still needs to rank the aggregate results.

That's not to say it won't have more significant impact in the future; I wouldn't know. But so far, I've yet to see the hype get realised.

Don't use it for things you're already an expert in, it can't compare to you yet.

Use it for learning new things, or for things you aren't very good at and don't want to bother with. For these it's incredible.

Perhaps.

> Do you remember when they teased GPT-2 as "too dangerous" for public access? I do. Yet we now have Llama 3 in the wild, which even at the smaller 8B size is about as powerful as the [edit: 6/13/23] GPT-4 release.

The statement was rather more prosaic and less surprising; are you sure it's OpenAI (rather than say all the AI fans and the press) who are hyping?

"""This decision, as well as our discussion of it, is an experiment: while we are not sure that it is the right decision today, we believe that the AI community will eventually need to tackle the issue of publication norms in a thoughtful way in certain research areas.

…

We are aware that some researchers have the technical capacity to reproduce and open source our results. We believe our release strategy limits the initial set of organizations who may choose to do this, and gives the AI community more time to have a discussion about the implications of such systems."""

I disagree here also: the company has openly acknowledged that this is a risk to be avoided with regards to safety related legislation, what they've called for looks a lot more like "we don't want a prisoner's dilemma that drives everyone to go fast at the expense of safety" rather than "we're good everyone else is bad".

People read too many sci-fi books and then project their fantasies on to real-world technologies. This stuff is incredibly powerful and will have social effects, but it’s not going to replace every single job by next year.

Have you tried asking it to generate a regex to transform your list into a CSV?

Now we have devs asking AI to generate regex formulas and pasting it into code without much concern on its validity.

Bad developers do bad regexes, regardless of whether they used AI.

kids these days and their lack of exposure to finite automata

- How so? I don't think it's possible to test for all cases...

- Well, it's easy, assuming a car on a non-branching track, moving with a constant speed and without any realistic external influences on it, you can simply calculate the distance traveled using the formula s = v/t. Ah, I wish I'd stop running into fools not knowing Newton's first law of motion...

- ??? Are you well?

I believe finite state modeling is used at NASA, A google search brings up a few references (that I'm probably not qualified to speak to), and I also remember hearing/reading a lecture on how they use them to make completely verifiable programs but can't find the exact one at the moment.

But not all regexes (eg, those using PCRE extensions afaik) are amenable to such a treatment. Those you just tend to hope they work.

The issue is with ChatGPT formatting a file.

All you have to do is survive long enough as an unemployed criminal until the system gets round to exonerating you:

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

"The British Post Office scandal, also called the Horizon IT scandal, involved Post Office Limited pursuing thousands of innocent subpostmasters for shortfalls in their accounts, which had in fact been caused by faults in Horizon, accounting software developed and maintained by Fujitsu. Between 1999 and 2015, more than 900 subpostmasters were convicted of theft, fraud and false accounting based on faulty Horizon data, with about 700 of these prosecutions carried out by the Post Office. Other subpostmasters were prosecuted but not convicted, forced to cover Horizon shortfalls with their own money, or had their contracts terminated. The court cases, criminal convictions, imprisonments, loss of livelihoods and homes, debts and bankruptcies, took a heavy toll on the victims and their families, leading to stress, illness, family breakdown, and at least four suicides. In 2024, Prime Minister Rishi Sunak described the scandal as one of the greatest miscarriages of justice in British history.

Although many subpostmasters had reported problems with the new software, and Fujitsu was aware that Horizon contained software bugs as early as 1999, the Post Office insisted that Horizon was robust and failed to disclose knowledge of the faults in the system during criminal and civil cases.

[...]

challenge their convictions in the courts and, in 2020, led to the government establishing an independent inquiry into the scandal. This was upgraded into a statutory public inquiry the following year. As of May 2024, the public inquiry is ongoing and the Metropolitan Police are investigating executives from the Post Office and its software provider, Fujitsu.

Courts began to quash convictions from December 2020. By February 2024, 100 of the subpostmasters' convictions had been overturned. Those wrongfully convicted became eligible for compensation, as did more than 2,750 subpostmasters who had been affected by the scandal but had not been convicted."

Even now, they're shipping text-image 4o but not the new voice while leaving old-voice up and confusing/disappointing a whole lot of people. This is a pretty big marketing blunder.

I remember for a good 2-3 months in 2023 ALL you could see on tiktok / youtube shorts was just garbage about 'how amazing' ChatGPT was. Like - video after video and I was surprised of the repeat content being recommended to me... No doubt openAI (or something) was behind that huge marketing push

I doubt OpenAI spent even one cent marketing their system (e.g. as in paying other companies to push it).

Who needs evidence when we have your lack of doubt hey?

It's ironic because the OpenAI creators really loved school and excelled academically. Nobody cares that ChatGPT destroyed advertising copywriting. But whatever little hope remained for the average high schooler post-lockdowns, it was destroyed by instant homework cheating via ChatGPT. So much for safety.

In a way I'm sorry, that's what people do nowadays. I'd prefer it to be paid for, honestly.

Job seekers currently in college have no idea what is about to hit them in 3-5 years.

In any other industry where just need an average margin of error close to a human's work and verification is much easier than generating possible outputs, the market will change drastically.

They won't "take 1 hour of time", they try it once or twice and give up.

"Intern of all trades, senior of none", to modernise the cliché.

Planning is different in that it is an essential part of agency. That's what Q* is supposed to add. My guess is that planning is the next type of functionality to be added to GPT. I wouldn't be surprised if they already have a version internally with such functionality, but that they've decided to hold it back for now for reasons such as safety (some may care about the election this year) or simply that the inference costs are so huge they cannot possibly expose it publicly.

Really philosophy seems to be one of the least important subjects right now. Hardly anyone learns about it in school.

If it was so important to success in the wild than it would stand to reason we all work hard at improving our reasoning skills, but very few do.

These did not provide useful life-lessons for me.

(The philosophy A-level I did voluntarily seemed to be 50% "can you find the flaws in this supposed proof of the existence of god?")

Shakespeare is packed with insight.

If I took out a loan on the value of goods being shipped to me, only for my ship to be lost at sea… it would be covered by insurance, and no bank would even consider acting like Shylock (nor have the motivation of being constantly tormented over religion) for such weird collateral, and the bank manager's daughters wouldn't get away with dressing up as lawyers (no chance their arguments would pass the sniff test today given the bar requirement) to argue against their dad… and they wouldn't need to because the collateral would be legally void anyway and rejected by any court.

The ships would also not then suddenly make a final act appearance to apologise for being late, to contradict the previous belief they were lost at sea, because we have radio now.

The closest to "relevant" that I would accept, is the extent to which some of the plots can be remade into e.g. The Lion King or Wyrd Sisters — but even then…

"Methinks, to employeth antiquated tongues doth render naught but confusion, tis less even than naughty, for such conceits doth veil true import with shadows."

We were naive to think we could have nice things for free.

It definitely took out a lot of jobs from the lowest rungs of the market, but on the more specialized / upper end of the ladder wages got actually higher and a lot of companies got burned, and now they have to readjust. It's rolling over slowly still, as there a lot of companies selling AI products and in turn new companies adopting those products. But it tells you a lot that

A) a company selling an AI assistant last year is now totally tied to automating busy work tasks around marketing and sales

B) AI writing companies are some of the busiest in employing human talent for... writing and editorial roles!

It's all very peculiar. I haven't seen anything like this in the past 15 years... maybe the financial crisis and big data was similar, but much much smaller at scale.

Effectively all mechanization, computerization, and I guess now AI-ization has done this. In the past you could have a rudimentary education and contribute to society. Then we started requiring more and more grade school, then higher education for years. Now we're talking about the student debt crisis!

At least if AI doesn't go ASI in the near term the question is how are we going to train the next generation of workers to go from unskilled to more skilled and useful than the AI is. Companies aren't going to want to do this. The individuals are going to think it's risky getting an education that could be replaced by a software update. If left to go out of control this is how a new generation of luddites will burn data centers in protest they are starving on the streets.

But OpenAI is having a hard time retaining/increasing ChatGPT users. Also, Alphabet's stock is about as valuable as it's ever been. So I don't think we have evidence that this is really challenging Google's search dominance.

But ChatGPT has really hurt Google's brand image.

The questions I still ask Google, have a lot of monetary value (restaurants, cloths, movie, etc).

But I agree seems SO often helps more than Google-AI.

Maybe not GPT-2, but in general LLMs and other generative AI types aren't without their downsides.

From companies looking to downsize their staff to replace them with software, to the work of artists/writers being devalued somewhat, to even easier scams and something like the rise of AI girlfriends, which has also gotten some critique, some of those can probably be a net negative.

Even when it's not pearl clutching over the advancements in technology and the social changes that arise, I do wonder how much my own development work will be devalued due to the somewhat lowered entry barrier into the industry and people looking for quick cash, same as with boot camps leading to more saturation. Probably not my position individually (not exactly entry level), but the market as a whole.

I spend a part of yesterday evening sorting my freshly dried t-shirts into 4 distinct piles. I used OpenAI Vision (through BeMyEyes) from my phone. I got a clear description of each and every piece of clothing, including print, colours and brand. I am blind BTW. But I guess you are right, no impact at all.

> Yet we now have Llama 3 in the wild

Yes, great, THANKS Meta, now the Scammers have something to work with. Thats a wonderful achievement which should be praised! </sarcasm>

That is a really great application of this tech. And definitely qualifies as real-world impact. Thanks for sharing that!

Go ask any teacher or graphician.

I can't help but notice the huge amount of hindsight and bad faith that it demonstrated here. Yes, now we are aware that the internet did not drown in a flood of bullshit (well, not noticeably more), when GPT-2 was released.

But was it obvious? I certainly thought that there was a chance that the amount of blog spam that could be generated effortlessly might just make internet search unusable. You are declaring "hype", when you could also say "very uncertain and conscientious". Is this not something we want people in charge to be careful with?

Even in this thread people talk about "Oh I use ChatGPT rather than Google search because Google is just stuffed with shit". And on HN there are plenty of discussions about huge portion of reddit threads being regurgitated older comments.

I feel like everyone who makes this claim doesn't actually have any data to backup it up.

Maybe that is GPT-5.

And this release really is just incremental improvements in speed, and tying together a few different existing features.

~8 years ago when self driving technology was all the rage and every major company was getting on board with ever more impressive technological demos, it seemed entirely reasonable to expect that we'd all be in a world of complete self driving imminently. I remember mocking somebody online around the time who was pursuing a class C/commercial trucking license. Yet now a decade later, there are more truckers than ever and the tech itself seems further away than ever before. And that's because most have now accepted that progress on such has basically stalled out in spite of absolutely monumental efforts at moving forward.

So long as LLMs regularly hallucinate, they're not going to be useful for much other than tasks that can accept relatively high rates of failure. And many of those generally creative domains are the ones LLMs are paradoxically the weakest in - like writing. Reading a book written by an LLM would be cruel and unusual punishment given then current state of the art. One domain I do see them completely taking over is search. They work excellently as natural language search engines, and "failure" in such is very poorly defined.

[1] - https://en.wikipedia.org/wiki/Sigmoid_function

I think what maybe seems not obvious amidst the hype is that there is a hell of a lot of engineering left to do. The fact that you can squash the weights of a neural net down to 3 bits per param and it still works -- is evidence that we have quite a way to go with maturing this technology. Multimodality, improvements to the UX of it, the human-computer interface part of it. Those are fundamental tech things, but they are foremost engineering problems. Getting latency down. Getting efficiency up. Designing the experience, then building it out.

25 years ago, early tech demos on the internet were promising that everyone would do their shopping, entertainment, socializing, etc... online. Breathless hype. 5 years after that, the whole thing crashed, but it never went away. People just needed time to figure out how to use it and what it was useful for, and discover its limitations. 10 years after that, engineering efforts were systematized and applied against the difficult problems that still remained. And now: look at where we are. It just took time.

It most definitely is not.

Let me know when you can get a Waymo to drive you from New York to Montreal in winter.

They are an existence proof that the original claim that we seem further than ever before is just wrong.

And level 4 is where Waymo is, and is staying. Their strategy is to to use tiny geofenced areas with a massive amount of preprocessing, mapping out every single part of an area, not just in terms of roads but also every single meta indicator - sign, signals, cross walks, lanes, and so on. And it creates a highly competent, but also highly rigid system. If road conditions change in any meaningful way, the most likely outcome with this strategy is simply that the network gets turned off until the preprocessing can be carried and reuploaded again. That's completely viable in small geofenced areas, but doesn't generalize at all.

So the presence of Waymo doesn't say much of anything about the presence of level 5 autonomy. If anything it suggests Waymo believes that level 5 autonomy is simply out of reach, because the overwhelming majority of tech that they're researching and developing would have no role whatsoever in level 5 automation. Tesla is still pushing for L5 automation, but if they don't achieve this then they'll probably just end up getting left behind by companies that double down on L4. And this does indeed seem to be the most likely scenario for the foreseeable future.

I don't think we have nearly as much visibility on what Waymo seems to believe about this tech as you seem to imply, nor do I think that their beliefs are necessarily authoritative. You seem disheartened that we haven't been able to solve self-driving in a couple of decades, and I'm of the opinion that geez, we basically have self-driving now and we started trying only a couple of decades ago.

How long after the invention of the transistor did we get personal computers? Maybe you just have unrealistic expectations of technological progress.

Another piece of evidence also comes from last year when Google scaled back Waymo with layoffs as well as "pausing" its efforts at developing self driving truck technology. [1] That technology would require something closer to L5 autonomy, because again - massive preprocessing is quite brittle and doesn't scale well at all. Other companies that were heavily investing in self-driving tech have done similarly. For instance Uber sold off its entire self-driving division in 2021. I'm certainly happy to hear any sort of counter-argument, but you need some logic instead of ironically being the one trying to mindread me or Waymo!

[1] - https://www.theverge.com/2023/7/26/23809237/waymo-via-autono...

15 years ago self driving of any sort was pure fantasy, yet here we are.

They'll release a version that can drive in poor weather and you'll complain that it can't drive in a tornado.

As long as the Waymo can be fed with the details, we'll be good. ;)

Joking aside, I think there are some cases where moving the goalposts is the right approach: once the previous goalposts are hit, we should be pushing towards the new goalposts. Goalposts as advancement, not derision.

I suppose the intent of a message matters, but as people complain about "well it only does X now, it can't do Y" - probably true, but hey, let's get it to Y, then Z, then... who knows what. Challenge accepted, as the worn-out saying goes.

This was 38 years ago: https://www.youtube.com/watch?v=ntIczNQKfjQ - "NavLab 1 (1986) : Carnegie Mellon : Robotics Institute History of Self-Driving Cars; NavLab or Navigation Laboratory was the first self-driving car with people riding on board. It was very slow, but for 1986 computing power, it was revolutionary. NavLab continued to lay the groundwork for Carnegie Mellon University's expertise in the field of autonomous vehicles."

This was 30+ years ago: https://www.youtube.com/watch?v=_HbVWm7wdmE - "Short video about Ernst Dickmanns VaMoR and VaMP projects - fully autonomous vehicles, which travelled thousands of miles autonomously on public roads in 1980s."

This was 29 years ago: https://www.youtube.com/watch?v=PAMVogK2TTk - "A South Korean professor [... Han Min-hong's] vehicle drove itself 300km (186 miles) all the way from Seoul to the southern port of Busan in 1995."

This was 19 years ago: https://www.youtube.com/watch?v=7a6GrKqOxeU - "DARPA Grand Challenge - 2005 Driverless Car Competition"

Meanwhile I've been using ChatGPT at work for _more than a year_ and it's been tremendously helpful to me.

This is not hype, this is not about how AI will change our lives in the future. It's there right here, right now.

The person I originally responded to stated, "We’re moving toward a world where every job will be modeled, and you’ll either be an AI owner, a model architect, an agent/hardware engineer, a technician, or just.. training data." And that far less likely than us achieving L5 self driving (if not only because driving is quite simple relative to many of the jobs he envisions AI taking over), yet L5 self driving seems as distant as ever as well.

But is this not what humans do, universally? We are certainly good at hiding it – and we are all good at coping with it – but my general sense when interacting with society is that there is a large amount of nonsense generated by humans that our systems must and do already have enormous flexibility for.

My sense is that's not an aspect of LLMs we should have any trouble with incorporating smoothly, just by adhering to the safety nets that we built in response to our own deficiencies.

Yep. So basically they're useful for a vast, immense range of tasks today.

Some things they're not suited for. For example, I've been working on a system to extract certain financial "facts" across SEC filings. ChatGPT has not been helpful at all either with designing or implementing (except to give some broad, obvious hints about things like regular expressions), nor would it be useful if it was used for the actual automation.

But for many, many other tasks -- like design, architecture, brainstorming, marketing, sales, summarisation, step by step thinking through all sorts of processes, it's extremely valuable today. My list of ChatGPT sessions is so long already and I can't imagine life without it now. Going back to Google and random Quora/StackOverflow answers laced with adtech everywhere...

The other day, I saw a demo from a startup (don't remember their name) that uses generative AI to perform financial analysis. The demo showed their AI-powered app basically performing a Google search for some companies, loosely interpreting those Google Stock Market Widgets that are presented in such searches, and then fetching recent news and summarizing them with AI, trying to extract some macro trends.

People were all hyped up about it, saying it will replace financial analysts in no time. From my point of view, that demo is orders of magnitude below the capacity of a single intern who receives the same task.

In short, I have the same perception as you. People are throwing generative AI into everything they can with high expectations, without doing any kind of basic homework to understand its strengths and weaknesses.

mapping th genome was that way. On a 20yr schedule, barely any progress for 15 and then poof, done ahead of schedule

My guess is the future belongs to those who don't stop—who, in fact, embrace the opposite of stopping.

I would even suggest that the present belongs to those who didn't stop. It may be too late for normal people to ever catch up by the time we realize the trick that was played on us.

Varying degrees of greedy / restless / hungry / thirsty / lustful are what we've got, because how is contentedness ever going to compete with that over millennia?

I've had a lot of negative things to say about religion for many years. However, as has been often observed, 'perception is reality' to a certain extent when it affects how people behave, and perhaps it's kind of a counterweight against our more selfish tendencies. I just wish we could do something like it without made up stories and bigotry. Secular humanist Unitarians might be about the best we can do right now in my opinion... I'm hoping that group continues to grow (they have been in recent years).

I have a much less "utopian" view about the future. I remember during the renaissance of neural networks (ca. 2010-15) it was said that "more data leads to better models", and that was at a time when researchers frowned upon the term Artificial Intelligence and would rather use Machine Learning. Fast forward a decade LLMs are very good synthetic data generators that try to mimic human generated input and I can't think somehow that this wasn't the sole initial intent of LLMs. And that's it for me. There's not much to hype and no intelligence at all.

What happens now is that human generated input becomes more valuable and every online platform (including minor ones) will have now some form of gatekeeping in place, rather sooner than later. Besides that a lot of work still can't be done in front of a computer in isolation and probably never will, and even if so, automation is not a means to an end. We still don't know how to measure a lot of things and much less how to capture everything as data vectors.

Currently the bottleneck is Agents. If you want a large language model to actually do anything you need an Agent. Agents so far need a human in the loop to keep them sane. Until that problem is solved most human jobs are still safe.

not quite sure that sanity is a business requirement

I fully expect GPT 5 (or at the latest 6) to similarly have native inclusion of agentic capabilities either this year or next year, assuming it doesn't already, but is just kept from the public.

Will be like, the end of millions of careers overnight.

It will probably strongly favour places like China and Russia though, where the economy is already strongly reliant on central control.

I think you may be literally right in the opposite sense to what I think you intended.

China (and maybe Russia) may be able to use central control to have an advantage when it comes to avoiding disasterous outcomes.

But when it comes to the rate of innovation, the US may have an advantage for the usual reasons. Less government intervention (due to lobbyism) combined with having several corporations actively competing with each other to be first/best usually leads to faster innovation. However, the downside may be the it also introduces a lot more risk.

The obvious way to to that is for it to plan a set of actions and evalute each possible way to reach some goal (or avoid an anti-goal). Kind of what AlphaZeros is doing for games. Q* is rumored to be a generalization of this.

This is no different to saying a person with a gun murdered someone rather than attributing the murder to the gun. An AI gun is just a really fancy gun.

A prompt is a _very_ different matter.

There may come a time where we grow so accustomed to this, that the decision is so heavily influenced by AI, that we believe it more than human decisions.

And then it can very well kill a human through misdiagnostic.

I think it is important to not just put this thought aside, but to evaluate all risks.

I would imagine outcomes would be scrutinized heavily for an application like this. There is a difference between a margin of error (existing with human doctors as well) and a sentient ai that has decided to kill, which is what it sounds like you're describing.

If we didn't give it that goal, how does it obtain it otherwise?

And “guns don’t kill people, people kill people”¹ is a bad argument created by the people who benefit from the proliferation of guns, so it’s very weird that you’re using that as if it were a valid argument. It isn’t. It’s baffling anyone still has to make this point: easy access and availability of guns makes them more likely to be used. A gun which does not exist is a gun which cannot be used by a person to murder another.

It’s also worth nothing the exact words of the person you’re responding to (emphasis mine):

> It can also murder people, and it will continue being used for that.

Being used. As in, they’re not saying that AI kills on its own, but that it’s used for it. Presumably by people. Which doesn’t contradict your point.

¹ https://en.wikipedia.org/wiki/Guns_don%27t_kill_people,_peop...

> I'd speculate that it's mostly people who don't use cars who criticize them the most, and the same with guns.

You mean that people who are opposed to something refuse to partake in its use and promotion? Shocker.

Everything always starts as a toy.

That includes, beyond literal Killers, all kinds of manufacturing, construction and service work.

I would expect a LOT of funds to go into research all sorts of actuators, artificial muscles and any other technology that will be useful in building better robots.

Companies that can get and maintain a lead in such technologies may reach a position similar to what US Steel had in the 19th century.

That could be the next nvidia.

I would not be at all surprised if we will have a robot in the house in 10 years that can clean and do the dishes, and that is built using basically the same parts as the robots that replace our soldiers and the police.

Who will ultimately control them, though?

Given the pace that AI is currently moving at, it seems to me that more and more, the mechanical aspect is becoming the limitation.

GPT 4o now seems to be quite good at reasoning about the world from pictures in real time. I would expect it would soon become easy for it to do the high level part of many practical tasks, from housekeeping to manufacturing or construction. (And of course military tasks.)

This leaves the direct low-level actuator control to execute such tasks in detail. But even there, development has been immense. See for instance these soccer playing robots [1]

And as both high level and low level control (if we assume that models soon will add agentic features directly into the neural networks), the only missing peace is the ability to build mechanically capable and reliable robots at a low enough price that they become cheaper than humans for various kinds of work.

There is one more limitation, of course, which is that GPT 4o still requires a constant connection to a data center, and that the models is too large to run within a device or machine.

This is also one of the most critical limitations of self driving. Had the AI within a Tesla had the same amount of compute available as GPT-4o, it should be massively more capable.

[1] https://www.youtube.com/watch?v=RbyQcCT6890

If you had an ASI? I don’t think you’d need a lot of funds to go into this area anymore ? Presumably it would all be solved overnight.

Companies that have a head start at that point, may get a huge first-mover advantage. Also, those companies also very well may have the capability to leverage AI in product development, just like everyone else.

And just as important as the products themselves is the manufacturing capacity to build them at scale. Until we have massive numbers of robots in service, building such infrastructure is likely to be slow and expensive.

EDIT: Also, once we really have the kind of Godlike ASI you envision, no human actions really matter (economically) anymore.

All AIs up to now lack autonomy. So I'd say until we crack this problem, it is not going to be able to do your job. Autonomy depends on a kind of data that is iterative, multi-turn, and learning from environments not from static datasets. We have the exact opposite, lots of non-iterative, off-policy (human made AI consumed) text.

So outside of use-cases where the user can quickly verify the result (like picking a decent generated image etc.),I can't see it being used much.

And guess what: RAG doesn't prevent hallucination. It can reduce it, and there are most certainly areas where it is incredibly useful (I should know, because that's what earns my paycheck), but it's useful despite still hallucinations being a thing, not because we solved that problem.

Regardless, we’ve seen accuracy of ~98% with simple context-based prompting across every category of generation task. Don’t take my word for it, a simple search would show the effectiveness of “n-shot” prompting. Framing it as “it _can_ reduce” hallucinations is disingenuous at best, there really is no debate about how well it works. We can disagree on whether 98% accuracy is a solution but again I’d assert that for >50% of all possible real world uses for an LLM 98% is acceptable and thus the problem can be colloquially referred to as solved.

If you’re placing the bar at 100% hallucination-free accuracy then I’ve got some bad news to tell you about the accuracy of the floating point operations we run the world on

I understand that you might be afraid. I believe that a world where only LLM companies rule the world is not practically achievable except in some distopian universe. The likelihood of the world where the only job are model architects, engineers or technicians is very very small.

Instead, let's consider the positive possibilities that LLMs can bring. It can lead to new and exciting opportunities across various fields. For instance, can serve as a tool to inspire new ideas for writers, artists, and musicians.

I think we are going towards a more collaborative era where computers and humans interact much more. Everything will be a remix :)

Oh, especially since it will be a priority to automate their jobs, or somehow optimize them with an algorithm because that's a self-reinforcing improvement scheme that would give you a huge edge.

GPT-4? Not that well. AI? Definitely

https://deepmind.google/discover/blog/alphageometry-an-olymp...

What's scary and cringey are your delusions.