Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

AI 2027

ai-2027.com

Post reply on HN
Print view View original

501–510 of 641 posts

Re: AI 2027

#501

Post by percentcer »

jug wrote: post #496
Vegenoid wrote: post #474

I think we've actually had capable AIs for long enough now to see that this kind of exponential advance to AGI in 2 years is extremely unlikely. The AI we have today isn't radically different from the AI we had in 2023. They are much better at the thing they are good at, and there are some new capabilities that are big, but they are still fundamentally next-token predictors. They still fail at larger scope longer ter…

> there are some new capabilities that are big, but they are still fundamentally next-token predictors Anthropic recently released research where they saw how when Claude attempted to compose poetry, it didn't simply predict token by token and "react" to when it thought it might need a rhyme and then looked at its context to think of something appropriate, but actually saw several tokens ahead and adjusted for where…

Isn't this just a form of next token prediction? i.e. you'll keep your options open for a potential rhyme if you select words that have many associated rhyming pairs, and you'll further keep your options open if you focus on broad topics over niche

Re: AI 2027

#502

Post by benlivengood »

Fraterkes wrote: post #484
benlivengood wrote: post #483

Earlier quoted context omitted.

METR [0] explicitly measures the progress on long term tasks; it's as steep a sigmoid as the other progress at the moment with no inflection yet. As others have pointed out in other threads RLHF has progressed beyond next-token prediction and modern models are modeling concepts [1]. [0] https://metr.org/blog/2025-03-19-measuring-ai-ability-to-com... [1] https://www.anthropic.com/news/tracing-thoughts-language-mod...

The METR graph proposes a 6 year trend, based largely on 4 datapoints before 2024. I get that it is hard to do analyses since were in uncharted territory, and I personally find a lot of the AI stuff impressive, but this just doesn't strike me as great statistics.

I agree that we don't have any good statistical models for this. If AI development were that predictable we'd likely already be past a singularity of some sort or in a very long winter just by reverse-engineering what makes the statistical model tick.

Re: AI 2027

#503

Post by kittikitti »

This is a great predictive piece, written in sci-fi narrative. I think a key part missing in all these predictions is neural architecture search. DeepSeek has shown that simply increasing compute capacity is not the only way to increase performance. AlexNet was also another case. While I do think more processing power is better, we will hit a wall where there is no more training data. I predict that in the near future we will have more processing power to train LLM's than the rate at which we produce data for the LLM. Synthetic data can only get you so far.

I also think that the future will not necessarily be better AI, but more accessible one's. There's an incredible amount of value in designing data centers that are more efficient. Historically, it's a good bet to assume that computing cost per FLOP will reduce as time goes on and this is also a safe bet as it relates to AI.

I think a common misconception with the future of AI is that it will be centralized with only a few companies or organization capable of operating them. Although tech like Apple Intelligence is half baked, we can already envision a future where the AI is running on our phones.

Re: AI 2027

#504

Post by 827a »

Philpax wrote: post #384
827a wrote: post #91

Readers should, charitably, interpret this as "the sequence of events which need to happen in order for OpenAI to justify the inflow of capital necessary to survive". Your daily vibe coding challenge: Get GPT-4o to output functional code which uses Google Vertex AI to generate a text embedding. If they can solve that one by July, then maybe we're on track for "curing all disease and aging, brain uploading, and coloni…

Haven't tested this (cbf setting up Google Cloud), but the output looks consistent with the docs it cites: https://chatgpt.com/share/67efd449-ce34-8003-bd37-9ec688a11b... You may consider using search to be cheating, but we do it, so why shouldn't LLMs?

I should have specified "nodejs", as that has been my most recent difficulty. The challenge, specifically, with that prompt is that Google has at least four nodejs libraries that are all seem at least reasonably capable of accessing text embedding models on vertex ai (@google-ai/generativelanguage, @google-cloud/vertexai, @google-cloud/aiplatform, and @google/genai), and they've also published breaking changes multiple times to all of them. So, in my experience, GPT not only will confuse methods from one of their libraries with the other, but will also sometimes hallucinate answers only applicable to older versions of the library, without understanding which version its giving code for. Once it has struggled enough, it'll sometimes just give up and tell you to use axios, but the APIs it recommends axios calls for are all their protobuf APIs; so I'm not even sure if that would work.

Search is totally reasonable, but in this case: Even Google's own documentation on these libraries is exceedingly bad. Nearly all the examples they give for them are for accessing the language models, not text embedding models; so GPT will also sometimes generate code that is perfectly correct for accessing one of the generative language models, but will swap e.g the "model: gemini-2.0" parameter for "model: text-embedding-005"; which also does not work.

Re: AI 2027

#505

Post by ComplexSystems »

Vegenoid wrote: post #474

I think we've actually had capable AIs for long enough now to see that this kind of exponential advance to AGI in 2 years is extremely unlikely. The AI we have today isn't radically different from the AI we had in 2023. They are much better at the thing they are good at, and there are some new capabilities that are big, but they are still fundamentally next-token predictors. They still fail at larger scope longer ter…

> They are much better at the thing they are good at, and there are some new capabilities that are big, but they are still fundamentally next-token predictors.

I don't really get this. Are you saying autoregressive LLMs won't qualify as AGI, by definition? What about diffusion models, like Mercury? Does it really matter how inference is done if the result is the same?

Re: AI 2027

#506

Post by Vegenoid »

gwd wrote: post #393
fmap wrote: post #366

Earlier quoted context omitted.

> even if this doesn’t lead to AGI, at the very least it’s likely the final “warning shot” we’ll get before it’s suddenly and irreversibly here. I agree that it's good science fiction, but this is still taking it too seriously. All of these "projections" are generalizing from fictional evidence - to borrow a term that's popular in communities that push these ideas. Long before we had deep learning there were people l…

> Someday we will have a machine simulate a cat, then the village idiot... This isn't how LLMs work. I think you misunderstood that argument. The simulate the brain thing isn't a "start from the beginning" argument, it's an "answer a common objection" argument. Back around 2000, when Nick Bostrom was talking about this sort of thing, computers were simply nowhere near powerful enough to come even close to being smart…

> Moore's law is exponential, which is where the "simulate a brain" predictions have come from.

To address only one thing out of your comment, Moore's law is not a law, it is a trend. It just gets called a law because it is fun. We know that there are physical limits to Moore's law. This gets into somewhat shaky territory, but it seems that current approaches to compute can't reach the density of compute power present in a human brain (or other creatures' brains). Moore's law won't get chips to be able to simulate a human brain, with the same amount of space and energy as a human brain. A new approach will be needed to go beyond simply packing more transistors onto a chip - this is analogous to my view that current AI technology is insufficient to do what human brains do, even when taken to their limit (which is significantly beyond where they're currently at).

Re: AI 2027

#507

Post by acureau »

eob wrote: post #413

An aspect of these self-improvement thought experiments that I’m willing to tentatively believe.. but want more resolution on, is the exact work involved in “improvement”. Eg today there’s billions of dollars being spent just to create and label more data, which is a global act of recruiting, training, organization, etc. When we imagine these models self improving, are we imagining them “just” inventing better math,…

This is exactly what makes this scenario so absurd to me. The authors don't even attempt to describe how any of this could realistically play out. They describe sequence models and RLAIF, then claim this approach "pays off" in 2026. The paper they link to is from 2022. RLAIF also does not expand the information encoded in the model, it is used to align the output with a set of guidelines. How could this lead to meaningful improvement in a model's ability to do bleeding-edge AI research? Why wouldn't that have happened already?

I don't understand how anyone takes this seriously. Speculation like this is not only useless, but disingenuous. Especially when it's sold as "informed by trend extrapolations, wargames, expert feedback, experience at OpenAI, and previous forecasting successes". This is complete fiction which, at best, is "inspired by" the real world. I question the motives of the authors.

Re: AI 2027

#508

Post by YetAnotherNick »

Aeolun wrote: post #358
YetAnotherNick wrote: post #353

Earlier quoted context omitted.

> very real risk of societal collapse or species extinction No, there is no risk of species extinction in the near future due to climate change and repeating the line will just further the divide and make the people not care about other people's and even real climate scientist's words.

Don’t say the things people don’t want to hear and everything will be fine? That sounds like the height of folly.

Don't say false things. Especially if it is political and there isn't any way to debate it.

Re: AI 2027

#509

Post by benlivengood »

Vegenoid wrote: post #491
benlivengood wrote: post #483

Earlier quoted context omitted.

METR [0] explicitly measures the progress on long term tasks; it's as steep a sigmoid as the other progress at the moment with no inflection yet. As others have pointed out in other threads RLHF has progressed beyond next-token prediction and modern models are modeling concepts [1]. [0] https://metr.org/blog/2025-03-19-measuring-ai-ability-to-com... [1] https://www.anthropic.com/news/tracing-thoughts-language-mod...

At the risk of coming off like a dolt and being super incorrect: I don't put much stock into these metrics when it comes to predicting AGI. Even if the trend of "length of task an AI can reliably do doubles every 7 months" continues, as they say that means we're years away from AI that can complete tasks that take humans weeks or months. I'm skeptical that the doubling trend will continue into that timescale, I think…

> I think there is a qualitative difference between tasks that take weeks or months and tasks that take minutes or hours, a difference that is not reflected by simple quantity.

I'd label that difference as long-term planning plus executive function, and wherever that overlaps with or includes delegation.

Most long-term projects are not done by a single human and so delegation almost always plays a big part. To delegate, tasks must be broken down in useful ways. To break down tasks a holistic model of the goal is needed where compartmentalization of components can be identified.

I think a lot of those individual elements are within reach of current model architectures but they are likely out of distribution. How many gantt charts and project plans and project manager meetings are in the pretraining datasets? My guess is few; rarely published internal artifacts. Books and articles touch on the concepts but I think the models learn best from the raw data; they can probably tell you very well all of the steps of good project management because the descriptions are all over the place. The actual doing of it is farther toward the tail of the distribution.

Re: AI 2027

#510

Post by 827a »

slaterbug wrote: post #209
827a wrote: post #91

Readers should, charitably, interpret this as "the sequence of events which need to happen in order for OpenAI to justify the inflow of capital necessary to survive". Your daily vibe coding challenge: Get GPT-4o to output functional code which uses Google Vertex AI to generate a text embedding. If they can solve that one by July, then maybe we're on track for "curing all disease and aging, brain uploading, and coloni…

You’ve intentionally hamstrung your test by choosing an inferior model though.

o1 fails at this, likely because it does not seem to have access to search, so it is operating on outdated information. It recommends the usage of methods that have been removed by Google in later versions of the library. This is also, to be fair, a mistake gpt-4o can make if you don't explicitly tell it to search.

o3-mini-high's output might work, but it isn't ideal: It immediately jumps to recommending avoiding all google cloud libraries and directly issuing a request to their API with fetch.

Post reply on HN