Microsoft and OpenAI's close partnership shows signs of fraying

But security and DLP teams will never accept it.

The Wikipedia part at least, is incorrect. Currently Wikipedia mods/admins are dealing with AI generated articles being uploaded.

As for NYT - I am assuming that lots of those stories are already available in some blog or the other.

The e-mail and web forums are 100% polluted with spam, which takes constant effort to remove. For GenAI based content, it is far harder to identify and remove.

This example assumes the effort required to keep the web functional can deal with AI created content. Speaking from experience, our filters (human and otherwise) cannot. They fail to do so even now.

PS: Even given your example of closed email chains - the information in that depends on sources people read. Like plastic pollution in the food chain, this is inescapable.

Twenty-four years later I still regret not being able to raise money to enable us to keep working on that nascent startup. In most ways it was still too early. Google was still burning through VC money at that point and the midwestern investors we had some access to didn't get it. And, honestly they were probably correct. Compute power was still too expensive and quality data sources like published text were mostly locked up and generally not available to harvest.

I'm asking because I know that with some prompts it gets the answer correct, and in those cases nothing in the tokenization has changed.

I can't use the OpenAI app anyway as they demand a logged-in Google account. I do have play services but not logged in.

It boggles my mind why they want to insist that you make an account with their total competition in order to use their service but they do.

Google and Meta aren't exactly lacking in conversation data: Facebook, Messenger, Instagram, Google Talk, Google Groups, Google Plus, Blogspot comments, Youtube Transcripts, &tc. The breadth and and breadth of data those 2 companies are sitting on that goes back for years is mind boggling.

If there's an actual business to be found in all this, that's where it's going to be.

The consumer side of this bleeds cash currently and I'm deeply skeptical of enough of the public being convinced to pay subscription fees high enough to cover running costs.

I have literally never heard that from anyone, and most everyone I know is “regular folk”.

I work in (large scale) construction, and no one has ever said anything even remotely similar. None of my non-technical or technical business contacts.

I’m not saying you haven’t, and that your in-group doesn’t, just that it’s not quite the cultural phenomenon you’re suggesting.

So when maintenance is required, it will be done by adding phrases like "Users should be able to select multiple receipts" to the existing script, and re-running it to regenerate the code from scratch.

Don't confuse the practical limitations of current models with conceptual ones. The latter exist, certainly, but they will either be overcome or worked around. People are just not as good at writing code as machines are, just as they are not as good at playing strategy games. The models will continue to improve, but we will not.

I can absolutely not go on the internet and download the source code for windows, because microsoft's security is impeccable.

With LLMs, if you have a use case which can run on an H100 or whatever and costs $4/hour, and the LLM has acceptable performance, it's going to be cheaper in a couple years.

Now, all these companies are improving their models but they're doing that in search of magical new applications the $4/hour model I'm using today can't do. If the $4/hour model works today, you don't have to worry about the cost going up. It will work at the same price or cheaper in the future.

I don’t think it’s that cut and dried though. Many users run into similar issues as other issues with things like reasoning (which is (allegedly) being addressed) and hallucinations (less so) both of which in turn become core reasons for subsequent better versions of the tech. Whether the subsequent versions deliver on those promises is irrelevant (though they often don’t) to that, at least IMHO, being a core reason to “stay on board” with the product. I have to think if they announced tomorrow they couldn’t afford to train the next one that there would be a pretty substantial attrition of paying users, which then makes it even harder to resume training in the future, no?

At some point it does make sense to have a small reactor powering a local datacenter or two, however. Licensing would still be not trivial.

From my view chatbots are still in the "selling dollars for 90 cents" category of product, of course it sells like discounted hotcakes...

OpenAI’s potential issue is that if Google offers tokens at a 10% gross margin, OpenAI won’t be able to offer api tokens at a positive gross margin at all. Their only chance really is building a big subscription business. No way they can compete with a hyperscaler on api cost long run

In case you are going to make an argument about how happiness or some related factor objectively determines success, let me head that off. Altman thinks that power rather than happiness determines success, and is also a human being. Why objectively is his opinion wrong and yours right? Both of your definitions just look like people's opinions to me.

The free speech part also reads completely hollow when the guy's first actions were to ban his critics on the platform and bring back self avowed nazis - you could argue one of those things are in favor of free speech, but generally doing both just implies you are into the nazi stuff.

SpaceX and Tesla have both accomplished great things. There's a lot of talented people that work there. Elon doesn'r deserve all the credit for all their hard work.

He’s dissecting it and connecting with the idea that if you a have a bigger vision and the ability to convince people, making a company is just an “implementation detail” … oh well .. you might be right after all … but I suspect is more nuanced, and is not endorsing religions as a means of obtaining success, I want to believe that he meant the visionary, bigger than yourself well intended view of it.

I think he is making an allusion to Apple's culture.

There's successful companies because their product is good, there's more successful companies because they started early (and it feels like a monopoly: Google, Microsoft), and there's the most successful company that tells you what you are going to buy (Apple's culture).

There's also mountains of research both theoretical and empirical that argue against exactly this point.

The problem is most papers on many scientific subjects are not replicable nowadays [0], hence my appeal to common sense, character, and wisdom. Highly underrated, especially on platforms like Hacker News where everything you say needs a double blind randomized controlled study.

This point^ should actually be a fundamental factor in how we determine truth nowadays. We must reduce our reliance on "the science" and go back to the scientific method of personal experimentation. Try lying to business partner a few times, let's see how that goes.

We can look at specific cases where it holds true- like in this case. There may be cases where it doesn't hold true. But your own experimentation will show it holds true more than not, which is why I'd bet against OpenAI

[0] https://en.wikipedia.org/wiki/Replication_crisis

I do not see how that is possible considering I have no clue who the second last owner of a cash was before me, most of the time.

What I remember from the rational optimist - with trust, trade is unlimited.

what I remember from debt - just too much, need to read it.

That said, why would an investor give money to altman if he is untrustworthy? it just gets worse and worse.

There is a difference between the two.

Models don't have this benefit. In Cursor, I can even switch between models. It would take a lot of convincing for me to switch off of Cursor, however.

It made the onboarding moderately easier for me.

Haven't successfully used any LLM at my day job though. Getting it to output the solution I already know I'll need is much slower then just doing it myself via auto complete

When I started using the LLM while coding, I was using Claude 3.5 Sonnet, but I was doing so with an IDE integration: Sourcegraph Cody. It was good, but had a large number of "meh" responses, especially in terms of autocomplete responses (they were typically useless outside of the very first parts of the suggestion).

I tried out Cursor, still with Claude 3.5 Sonnet, and the difference is night and day. The autocomplete responses with Cursor have been dramatically superior to what I was getting before... enough so that I switched despite the fact that Cursor is a VS Code fork and that there's no support outside of their VS Code fork (with Cody, I was using it in VS Code and Intellij products). Also Cursor is around twice the cost of Cody.

I'm not sure what the difference is... all of this is very much black box magic to me outside of the hand-waviest of explanations... but I have to expect that Cursor is providing more context to the autocomplete integration. I have to imagine that this contributes to the much higher (proportionately speaking) price point.

On the other hand if you mean, give you the correct answer to your question 100% of the time, then I agree, though then what about things that are only in your mind (guess the number I'm thinking type problems)?

Either the next tokens can include "this question can't be answered", "I don't know" and the likes, in which case there is no omniscience.

Or the next tokens must contain answers that do not go on the meta level, but only pick one of the potential direct answers to a question. Then the halting problem will prevent finite time omniscience (which is, from the perspective of finite beings all omniscience).

Liken them to climate-deniers or whatever your flavor of "anti-Kool-aid" is

https://news.ycombinator.com/newsguidelines.html

I'm not a huge fan of AI, but even I've seen articles written about its limitations.

Here's a great example:

https://decrypt.co/126122/meet-chaos-gpt-ai-tool-destroy-hum...

Sooner than even the most pessimistic among us have expected, a new, evil artificial intelligence bent on destroying humankind has arrived.

Known as Chaos-GPT, the autonomous implementation of ChatGPT is being touted as "empowering GPT with Internet and Memory to Destroy Humanity."

So how will it do that?

Each of its objectives has a well-structured plan. To destroy humanity, Chaos-GPT decided to search Google for weapons of mass destruction in order to obtain one. The results showed that the 58-megaton “Tsar bomb”—3,333 times more powerful than the Hiroshima bomb—was the best option, so it saved the result for later consideration.

It should be noted that unless Chaos-GPT knows something we don’t know, the Tsar bomb was a once-and-done Russian experiment and was never productized (if that’s what we’d call the manufacture of atomic weapons.)

There's a LOT of things AI simply doesn't have the power to do and there is some humorous irony to the rest of the article about how knowing something is completely different than having the resources and ability to carry it out.

I don't know if GPT-4 is smart enough to be successful at something like what OP describes, but I'm pretty sure it could cause a lot of trouble before it fails either way.

The real question here is why this is concerning, given that you can - and we already do - have humans who are doing this kind of stuff, in many cases, with considerable success. You don't need an AI to run a cult or a terrorist movement, and there's nothing about it that makes it intrinsically better at it.

They can certainly appear to be very smart due to having the subjective (if you can call it that) experience of 2.5 million years of non-stop reading.

That's interesting, useful, and is both an economic and potential security risk all by itself.

But people keep putting these things through IQ tests; as there's always a question about "but did they memorise the answers?", I think we need to consider the lowest score result to be the highest that they might have.

At first glance they can look like the first graph, with o1 having an IQ score of 120; I think the actual intelligence, as in how well it can handle genuinely novel scenarios in the context window, are upper-bounded by the final graph, where it's more like 97:

https://www.maximumtruth.org/p/massive-breakthrough-in-ai-in...

So, with your comment, I'd say the key word is: "currently".

Correct… for now.

But also:

> All these chatgpt things have a very limited working memory and can't act without a query.

It's easy to hook them up to a RAG, the "limited" working memory is longer than most human's daily cycle, and people already do put them into a loop and let them run off unsupervised despite being told this is unwise.

I've been to a talk where someone let one of them respond autonomously in his own (cloned) voice just so people would stop annoying him with long voice messages, and the other people didn't notice he'd replaced himself with an LLM.

If you ask it to make a plan, it will spit out a sequence of characters reasonably indistinguishable from a human-made plan. Sure, it isn’t “planning” in the strict sense of organizing things consciously (whatever that actually means), but it can produce sequences of text that convey a plan, and it can produce sequences of text that mimic reasoning about a plan. Going into the semantics is pointless, imo the artificial part of AI/AGI means that it should never be expected to follow the same process as biological consciousness, just arrive at the same results.

It’s true that understanding is quite primitive at the moment, and it will likely take further breakthroughs to crack long horizon problems, but even when we get there it will never understand things in the exact way a human does. But I don’t think that’s the point.

This is super incorrect. The base model is trained to predict the distribution of next words (which obviously necessitates a ton of understanding about the language)

Then there's the RLHF step, which teaches the model about what humans want to see

But o1 (which is one of these LLMs) is trained entirely differently to do reinforcement learning on problem solving (we think), so it's a pretty different paradigm. I could see o1 planning very well

We're already starting to see signs of that even with GPT-3, which really was auto-complete: https://academic.oup.com/pnasnexus/article/3/2/pgae034/76109...

Fortunately even the best LLMs are not yet all that competent with anything involving long-term planning, because remember too that "megalomaniac" includes Putin, Stalin, Chairman Mao, Pol Pot etc., and we really don't want the conversation to be:

"Good news! We accidentally made CyberMao!"

"Why's that good news?"

"We were worried we might accidentally make CyberSatan."

I don’t consider myself an AI doomer by any means, but I also don’t find arguments of the flavor “it just predicts the next word, no need to worry” to be convincing. It’s not like Hitler had Einstein level intellect (and it’s also not clear that these systems won’t be able to reach Einstein level intellect in the future either.) Similarly, Covid certainly does not have consciousness but was dangerous. And a chimpanzee that is billions of times more sophisticated than usual chimps would be concerning. Things don’t have to be exactly like us to pose a threat.

It's true, they might not think the way we do.

But reasoning can be formulaic. It doesn't have to be the inspired thinking we attribute to humans.

I'm curious how you define "reason".

As a mere software engineer who's made a few (pre-transformer) AI models, I can't tell you what "actual cognition" is in a way that differentiates from "here's a huge bunch of mystery linear algebra that was loosely inspired by a toy model of how neurons work".

I also can't tell you if qualia is or isn't necessary for "actual cognition".

(And that's despite that LLMs are definitely not thinking like humans, due to being in the order of at least a thousand times less complex by parameter count; I'd agree that if there is something that it's like to be an LLM, 'human' isn't it, and their responses make a lot more sense if you model them as literal morons that spent 2.5 million years reading the internet than as even a normal human with Wikipedia search).

But sure, if you have an un-embodied super-human AGI you should assume that it can figure out a super-human shelf-stocking robot shortly thereafter. We have Atlas already.

Physical embodied (generally low-skill, low-wage) work like cleaning and carrying things is likely to be some of the last work to be automated, because humans are likely to be cheaper than generally capable robots for a while.