i'm not saying that LLMs won't keep getting marginally better, and that people will find very useful ways of implementing them in practice - they will. but any talk of "AGI" or generalized capabilities in LLMs is at this point knowingly and deliberately misleading and meant to deceive the public and investors

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462237)

As big a problem as (maybe bigger than) the technologists themselves is VC. The firms VCs invest in aren't publicly traded, so there's no opportunity to short, resulting in a field where no one shits on anyone else's investment and it's a pure one-way positive feedback loop. Venture capitalists are known to get catty and defensive as fuck at the most gentle/accidental passing criticism from other VCs, because that's not how the game is played and everyone should know that a rising tide of hype raises all fraud ships.

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462432)

meanwhile actual american businesses in industries that actually make and do important stuff can't get any funding at all

"private equity" sure seems like nothing more than an entire "industry" of pump and dump scamming

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462450)

I am by no means some deep ideological believer in crypto, but when TWTR reached a market cap of $60B while posting growing YOY losses as their -4% profit margin widened--with no plan for doing anything different; dorsey would just give different answers everytime he was asked, and that other turdskin would just start babbling in hindi or whatever and hope that the white-girl CNBC reporter would start to wonder if she picked out the wrong turd from the conference room and just abashedly back away--I really started to appreciate the comparatively Sound Fundamentals of the BTC fake (and never actually used) internet-money market.

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462485)

and then it turns out its value prop the whole time was being a shiny object to the world's richest sperg

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462624)

The only other thing I would add is that VC switched from a risk sharing model (1970s to around 2008) where partners at the firms would contribute their own capital; to a risk shofting model (2009-Present) where VC funds mostly gather assets and then stuff junk into funds and charge a fee.

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462654)

it won't happen overnight though. our Gen is probably still largely fine, maybe we'll have to maek it as policemen at 50 or something. but anybody under 30 and especially those coming out of college now are going to experience a much different job market

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462481)

btw i was completely right about models being specifically RL trained to perform on evals, and that's the reason why they're performing increasingly well on evals while overall capabilities are stagnating (duh)

machine learning pumo here didn't believe me, but now we've got the leading guys in the industry admitting it outright

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462562)

what does this even mean? I don't think anybody ever trained an LLM on Pathfinder 1E rules but it does a decent job of sorting through something that is comically in depth.

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462645)

LLMs can only produce stuff that's within its training distribution. everything outside of its training is Inaccessible for the LLM

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462653)

for example, in l*w: LLMs do a pretty good job, actually, of evaluating and "reasoning" about cases that have the same structure as other cases in their training data. but any remotely novel kind of legal case, and they totally fail

humans *can* generalize, at least to a far greater extent than LLMs. we can see a completely novel legal case and reason about it pretty well. we can create a mental model of the whole case and think about it accurately and coherently. LLMs can't do that at all

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462704)

they just can't do this, you can try it yourself

making shit up that doesn't work and doesn't make sense is not "generalization"

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462760)

If you're trying to do something truly novel LLMs struggle, but this is like <1% of all knowledge work. To be fair law probably lags behind other fields in data training just because a lot of stuff is just locked behind privatized paywalls, but this will inevitably change over time.

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462801)

they just can't generalize, it's as simple as that. it's why AI implementation is lagging so much. if LLMs were really capable of generalization, they could be plug and played into virtually any role. but they can't, so they aren't

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462885)

And law, unlike programming for the other 10% who are beyond crudmonkeys like me, doesn't involve creating new stuff -- it is by definition the application and analysis of a set of existing rules. Knowledge professions will die off because there's just rapidly declining value in the knowledge itself, you'll be left with a fraction of the current workforce which basically audits and lightly curates AI generated output.

You should not mistake what people like Karpathy or Ilya are talking about with the replacement of 90% of white collar knowledge profession BS. That's considered the low hanging fruit on the path to AGI, the tricky part as far as that goes is getting rid of cartels, building of trust, and building out the software pipelines to do so. But considering how much companies stand to gain it is a problem that will be solved.

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49469789)

Are you maybe a few crucial years younger than me? Were you not around for Nanotechnology?

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462854)

Imagine if The Music Man got to shill techno-utopia as a replacement for the fent epidemic, instead of shitty brass-band instruments as a replacement for pool halls. That's what we're dealing with here.

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462909)

that's retarded. that's like the people who said "but there's no money in search lmao" 30 years ago and google now prints dizzying sums of money.

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462823)

.

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462741)

NotebookLM

AI call notetaking (Otter or similar)

Claude/ChatGPT/Gemini projects for doing first drafts of stuff where I simply want to drop in voluminious notes/documents and let it do what it do

More broadly, AI is going to be game-changing in may fields, particularly science. It analyzes the literature on a subject and analyzes experimental data much more quickly than previous methods. I'm also told it can design novel drugs/enzymes/ etc. I expect the pace of scientific advances to increase, with unpredictable results.

I expect the capabilities of AI within current use cases will get better, and that new use cases will emerge. I still think it's going to eliminate the need for a lot of human mental work, with unpredictable consequences

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462743)

one of my favorite all time xo poasts I lost track of many years ago was a pretty big thread where everyone shit on the outrageous sum of money facebook spent acquiring instagram. it was 1 billion dollars and instagram would go on to be the very reason facebook thrived even as its core platform became entirely boomers. in fact primary thing facebook was actually good at was picking which companies to buy early on (instagram and whatsapp have been insane returns on investment)

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462833)

Notice the leaps these charlatans make in describing their product.

AGI (like the stat in warcraft) used to be AI used to be Machine 'Learning' used to be Linear Regression.

Name ONE (uno) (1) novel invention or one nontrivial scientific discovery that is for the benefit of humanity and is not sophistry.

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462865)

https://www.youtube.com/watch?v=Fk0QC7wf2do

heres your average tech evangelist

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462943)

Just comparing to ecommerce there were these secondary and tertiary problems that had to be properly solved re logistics infrastructure, global supply chains, payment systems, digital marketing and advertising, compute infrastructure etc. that spurred all sorts of massive innovations and spawned entirely new lines of business that were even more profitable and strategic, and that’s before you even get into the massive benefit that accrued to a company like Amazon when the iPhone came along/was so good and mobile quickly became the primary use case. We’re in the 99-00 phase of the investment hype cycle but AI companies are already working out various solutions to eg the energy demand problem as a macro limitation, but it’s problems along those lines that need to be credibly and durably solved to unlock adoption and implementation levels that will actually justify the current insane levels of investment.

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49462867)

It's certainly useful but it still seems a ways off... Right now it's primarily feedback loop is to validate the user and there's a lot of limitations with the paid versions in terms of what it will generate, how it generates it, etc.

For ex. just creating a basic manual regarding expanding into a new practice area required hours of prompting and tweaking and dealing with the website's bullshit to get it to work. Tried this with paid version of both claude and chatgpt and each had their own specific limitations.

But the bigger concern is the constant validation of whatever the user is feeding it since their whole goal is to drive engagement. In some ways it dumbs down the product.

(http://www.autoadmit.com/thread.php?thread_id=5802800&forum_id=2)#49463841)