The AI Market Is Firming Up Fast | Hacker News Reader

The AI Market Is Firming Up Fast

dbreunig.com

48 pointsdbreunig20 comments

There may also be alternative solutions to the problem solving that ML offers we just stopped thinking about them because so much r&d goes into neural networks.

The definition of AI used in the executive order ([1]) does not include neural nets or even ML / statistics, only the ability to make predictions/recommendations/decisions. It is not apparent what kind of R&D in this area would not fall under it:

>(b) The term “artificial intelligence” or “AI” has the meaning set forth in 15 U.S.C. 9401(3): a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments. Artificial intelligence systems use machine- and human-based inputs to perceive real and virtual environments; abstract such perceptions into models through analysis in an automated manner; and use model inference to formulate options for information or action.

1. https://www.whitehouse.gov/briefing-room/presidential-action...

I don’t think this correct, the key being here this sentence:

> abstract such perceptions into models through analysis in an automated manner

Companies aren't worried about "problem solving" they're worried about "using AI".

Not long ago I was at a company that was rushing to get their AI "chat with our documentation" bot out the door as fast as possible. But prior to AI hype, 'improve documentation search' was not even remotely a priority, or even a concern. It's not like they had an okay chatbot for docs that was previously important but not high enough quality. Why was a documentation chatbot suddenly important?

I was talking to a neighbor the other day and he said he was tasked with using AI to help solve scheduling problems in his hospital! This was particularly bizarre, because if "solve scheduling issues" was a real problem, there are plenty of existing approaches to this (and have been for decades) and further more LLMs are not a particularly good way to solve this.

There are plenty of tricky NLP problems that AI/LLMs can unblock, and it makes sense that if you've been trying to solve a tricky NLP problem you might think AI can help you. There are numerous old problems I've had that I just couldn't get quite performant enough to work years ago that I would love to revisit with LLMs.

Funnily enough, when I chat with former coworkers about progress they've made solving these old problems they tell me leadership isn't interested and instead they're building something like a chatbot for documentation.

Really nice to hear someone write this out. I have a similar perspective about how companies are just trying to have _any ol'_ AI offerings for the sake of "not being left behind" or something. But in practice they don't offer much.

At work there are meetings promoting AI tooling like "Meet ChatGPT your new best friend" and how to leverage ai tools internally - seems like just another thing to jump onto after NFTs

There is a similar thought leader bullshit hype cycle to NFTs. There is also something fundamentally different about LLMs in that they are actually useful. I hosted a “how to use GitHub Copilot” session at work. It was all based on the positive experience of using Copilot to do my normal job. It wasn’t something imagined for someone else or in the future.

There was a Copilot session yesterday. I respect your opinion and agree it can be used in a positive manner. Tough to cut through all the b.s. though and I'm incredibly wary

I'm pretty sure every bigco is currently putting a bunch of money (a bunch) into the probably-false promise that putting an LLM in front of their data will finally give their execs the excellent near-real-time insight into their entire business that they've been burning cash trying and failing to achieve since like the 80s.

New flavor of the same shit-filled ice cream. Don't matter what flavor it is. There's shit in it.

But execs just cannot help themselves. The dream of getting perfect visibility into their whole org from a dashboard (this time, with AI chat prompt!) on their phone, that they can check on a whim from the 14th hole, is just too enticing to let it go as the false promise that it always is. The mind boggles at how much money has been wasted on this kind of garbage over the decades.

Probably the AI sorting out documentation seems cheaper than the effort it would take previously to improve it.

There are a lot of benefits to the constitutional prohibition on ex-post-facto laws but one of the downsides is the way the private sphere ends up lead by the fastest-drawing outlaws.

Regarding data, an idea would be if data had a time moratorium, where it had a do-not-use until the data was x days old.

That's how copyright works, but the courts have sided against your copyright in this area, it would take a new law to give it back. Congress needs to say, "this precedent is wrong, you have a copyright to your data and ChatGPT is a massive infringer, way worse than the internet archive would be if it gave away free books."

If the world was perfect we would reduce copyright terms to twenty to thirty years and allow training AIs on public domain text... solving the human learning (free books relevant to today!) and AI learning issues in one step. Sadly the people on the inside want to take everything and leave us with nothing so we will probably get a hodgepodge of laws, a philosophical system no one can explain, and a seriously nonuniform landscape of compromises. Don't interpret this as defeatism, just as pointing out where the ball is heading if nobody catches it.

If you want one takeaway from my comment, don't let them train AI on text you are not allowed, yourself, to read. See this lawsuit for evidence that it is happening: https://www.cbc.ca/radio/asithappens/authors-guild-chatgpt-l...

The market for using multi head attention based transformers is firming up fast, but it's one algorithmic innovation away from all those foundation models being worthless and starting over from scratch.

>allowing them to be built with less compute and less data – enabling more challengers to enter and create a more dynamic field during these formative years.

I feel the solution will never be using less compute and less data, but rather, will those projects like Gensyn and Bittensor actually succeed in the sharing of compute and data resources.

I dunno anything about this crypto, but just getting community ML projects to integrate basic optimizations and hw vendor ports (much less rearchitect for these projects) is like pulling teeth. If its not turnkey PyTorch... Yeah.

Also, Petals and AI Horde seem to have gotten far even though they are relatively new (and crypto free)

>Further, those parties make up an ouroboros of funding, where invested funds quickly come back to investors in the form of cloud compute fees. This allows Amazon, Microsoft, and Google to stomach sky-high valuations for these rounds, as they’re essentially buying marketshare for their cloud offerings. This dynamic is a major barrier standing between you and the latest GPUs.

I don't think cloud compute prices are so far above their costs that this would amount to a huge distinguishing factor.

But equities investors value cloud growth for all of these companies, and AI spend is high enough that even if the margin isn't there, the revenue growth will make the street happy.

But isn't that the same kind of sales growth an investor could see if a retail company used investor money to load up trucks with its own shelved goods?

Not necessarily compute minutes but ancillary costs like egress and especially local block storage. I recently tested a training cluster and these costs were nearly as much as the GPUs. And they’re both extremely high margin as well as lower QoS than what you can get in bare metal.

There is not going to be a chance for small folks in AI land. It looks more like military industry day by day.

Reply on news.ycombinator.com