Haven't they already proven to be extremely useful? In some areas they are definitely here to stay, coding/software and search (retrieve and summarize information). There's a bunch of places where they are surely shoehorned in, overhyped, and don't belong, but there's also equally many places where they might still be transformative but aren't used yet.
But overall I think the technology is well proven.
I always leave room open for failure, and that approach has generally served me well personally and professionally. I have never been punished for having an exit strategy.
Besides, the marketplace is still in its infancy for LLMs, with a lot of unanswered questions. A lot of those questions surround the commercial viability of frontier models on bespoke hyperscaler data centers with limited usage outside of LLMs specifically should those economics be non-viable. Since that's where the memory is being tied up into, that means it's a critical question to answer in order to determine long-term investment needs into further memory fabrication.
I got an RTX 6000 pro too. I like running locally, I've learned a lot more than if I had used an API and there's less worry about overspending tokens. I accidentally spent $100 on claude api in like 2 days because I didn't know what I was doing.
The problem is that while one these gpus is a huge improvement over a laptop or a single 3090, you very quickly wish you had more. I would buy a second one, but I did the math and realized that with the current crop of models, 2 Blackwells doesn't buy me any new capability that I didn't have with one. So I would need a 3rd one. And when I buy a 3rd one I will feel like I want to running a higher quant, so then I will want a 4th.
A pair of RTX6000 cards will give you a good performance boost due to tensor parallelism, though. I haven't tried the newest predictive quants but I see about 35 tps when running the 8-bit Qwen 3.6 27B model on one board and about 50 tps on two. Probably could come close to 100 tps on an optimized setup with the latest GGUFs.
Also, the 4-bit quants of MiniMax 2.7 will run at 100 tps or so with two cards, which is pretty decent. It doesn't go any faster at all with 4 GPUs from what I've seen, so if you don't actively need 384 GB of VRAM, 2x RTX6000 is a good place to be.
Using an Epyc platform to get plenty of PCIe lanes and memory channels. I have couple of extra 3090s plugged in which get some offload and help with larger models that don't fit entirely on the blackwell.
At this point IPOs are mainly for unloading bags onto retail. Every institution who wanted a piece of these labs got in years ago and captured all the value.
Well, sad to say this is simply untrue for a few reasons.
1. "Retail" does not have enough purchasing power to have all of these "bags" unloaded on to.
2. Institutions buy shares in public firms post-IPO all the time even when they're "unloading bags onto retail". Take Uber (random example) ~83% is owned by institutions.
3. General factual history of the stock market shows that you are incorrect. Successful companies that IPO and continue to do business still have quite a lot of room left to grow. What was Google's market capitalization at IPO? What is it now? Is it possible some early investors made higher multiples than the IPO -> May 20th valuation? Yea for sure. That doesn't mean that all the value was captured. It also doesn't take into account the early stage risk for investing. Is Google an "at this point IPO"? No, but the principle is the same.
It's also worth mentioning however that the number of IPOs is going down over time. You could maybe argue that the only ones that actually IPO are all the bags, but that seems like a stretch.
These cynical comments "IPOs are mainly for unloading bags on to retail" lack explanatory power and data.
It's absolutely true. Just look at how private equity is now getting access to public markets and retirement accounts[0]. You think PE is letting the little guys in out of the goodness of their hearts? No, they've extracted as much as they can and the market is starting to question the absurd valuation of private assets.
A wise man once said: "if you're given an opportunity to cut an amazing deal and you can't tell who's getting screwed, then it's probably you"
So I take it you're going to buy shares of OpenAI on opening day then? ;)
Institutions merely owning a newly-IPO'd stock means nothing. They get access to shares at a reasonable price before opening while retail is buying at insane prices after open. See Figma as an example where institutional investors got it at $33/share and it ended the IPO day at $115/share with retail buying all the way up (including pops above that at like $127)
I thought it was common knowledge that IPOs are a way for insiders and early investors (not IPO flippers) to get a nice exit during the frenzy.
> So I take it you're going to buy shares of OpenAI on opening day then? ;)
Probably not. Do you understand however that your comment does not make sense in the context of my comment?
> Institutions merely owning a newly-IPO'd stock means nothing. They get access to shares at a reasonable price before opening while retail is buying at insane prices after open. See Figma as an example where institutional investors got it at $33/share and it ended the IPO day at $115/share with retail buying all the way up (including pops above that at like $127)
It also doesn't mean nothing - you have to go and analyze any given stock to make these kinds of claims on a per-IPO/equity basis. You also are ignoring traders and trading algorithms run by... big institutions and trading firms, and you're not accounting for volume or accounting for post-IPO purchases nor breaking those down by segment. In other words, you're just making stuff up.
I understand that, but what I'm not understanding is why this seems to be a concern. I suppose equity given to early employees is a problem too and they're just "dumping their bags on retail" after their lockup period expires?
Earlier stage investors take risk and are rewarded for that. Most companies go bankrupt and folks lose their principal. For the companies that are successful yea some go bust after IPO - so what? Are you against public markets or something? That would at least be an interesting discussion.
Google IPO'd in 2004 and returned from what I'm reading about 6,500% after IPO (and this was in 2024, so the gains have gone up much higher since then) and all of that was the bags dumped on retail. If someone wants to dump their 6,500% return on me I'll take them up on that all day every day and twice on Sunday.
I use claude/gemini as my homepage now (I have to keep switching as these companies make "updates" that periodically render their models useless). Even if I want to search for simple things, I would rather have an LLM wade through the result and extract just the information I asked for. SEO, and now mountains of slop content have made this necessary. Only a matter of time before the SEO industry in large figures out how to game LLMs too, making them equally useless.
I already saw a article recently about how to set up a business domain which can reliably show up in a search result and dump overly positive reviews into anyone's context.
They won't, its literally part of their sales funnel. They've specifically engineered a bad experience for anyone outside the ecosystem by making it all of their friend's problem too. Its very important for their stock price that text messages sent by non apple products are just slightly more difficult to read.
Is this LLM psychosis? So much tending and conversing with the matmuls but what was the outcome? Are people who get this into it more successful somehow? It reminds me of people who take drugs and get "revelations" but then are not particularly over represented in the group of successful people for all of their deep insights.
> It reminds me of people who take drugs and get "revelations" but then are not particularly over represented in the group of successful people for all of their deep insights.
This depends on where you're looking for "successful" people.
I generally agree with you - of those people who might report "revelations" through hallucinogenic drugs, the majority may misinterpret their drug-induced experience and hence be more confused / lost than before.
On the other hand, it can still be true that among those who eventually do have genuine spiritual insight, having used hallucinogenic substances is overrepresented compared to the general population.
Quoting from [1], where the author tried to find spiritually advanced individuals:
> Approximately 52% of
participants had used hallucinogenic drugs at
some point; none reported these as the trigger
that led to PNSE.
PNSE = Persistent Non-Symbolic Experience.
My point is: while there are certainly people who go way overboard with the LLM stuff, that is not at odds with skillful use of LLMs being overrepresented in successful people.
I see now that you didn't make that point, but I already typed this all out and I'm gonna leave it.
LLMs vs psychedelics are an interesting comparison: psychedelics lower the threshold for the epiphany sensation, arguably LLMs do the same through agreement (‘sycophancy’).
With psychedelics you see the upsides in engineers who come to work the next day with designs informed by massive insights, and the downsides in people whose ‘massive insights’ result in socially abnormal behaviours.
LLMs seemingly provide validation and support, something very lacking in most lives. The distance between thinking loosely about, say, an app and getting positive feedback and absurd market predictions triggers similar ‘revelations’ in people with no meaningful context, and plays on the hopes of those who do. Much like with psychedelics, though, there is a self-supported self-validated cycle going on, prone to flights of excess.
Coming out publicly about doing drugs is just not something that smart people usually do.
There's a lot of stigma about it and a lot of opinionated people who don't know anything about them and still have a lot of judgements.
Then there are these people calling psychedelics hallucinogens. The former can give spiritual revelations and the latter gives vividly realistic, usually nightmarish, hallucinations.
There are also people putting an equal sign between crack cocaine and psylocibine mushrooms. But I don't think an addict will rob a petrol station with a screwdriver just to get another dose of magic mushrooms.
In such surroundings, it's just not a good idea to publicly say anything about drugs, unless you're so rich or known for your expertise, that you can shoulder some negative judgement and it's implications.
Psychosis would be a mental illness with symptoms like losing track of reality. I think there might actually be a bit of an anti-AI psychosis where individuals are projecting their fear and paranoia onto others and seeking confirmation bias. The need to dismiss the success others are having with AI could be triggered by personal insecurities. And of course deep paranoia is one of the classical symptoms of an actual psychosis. Just holding up a mirror here.
BTW. I have friends and relatives that have dealt with an actual psychosis. Not fun to experience up close. So, I don't want to take this metaphor too far. But if we're using big words like psychosis here, you might want to examine your motives, insecurities and reasoning a bit.
I've played with agentic AIs for coding and other use cases and have had some successes and failures. I'm fairly impressed with some stuff that is possible now and I use technology pragmatically. As I always have throughout my 30 year career.
For any new thing, there are always early adopters and those who really don't get it. And most other people let the early adopters figure things out and then end up copying what works some months/years later. And you always have some stragglers that can't or won't adapt that can't be helped.
Most of what this person describes in the article is very reasonable. Codex might not be the best model. But in terms of UX, OpenAI is getting a few things right that starting to make a difference. It's only a few months ago that their desktop app launched. It has gone through a pretty rapid evolution. As of a few weeks ago (it's that recent) you can install skills to connect your gmail, canva, google drive, work with ppt files, etc. In other words, it's now suitable for things other than programming. Before that, Claude CoWork was a few weeks earlier. So, this is all very new and fresh. I've tried some of this stuff and it mostly works as advertised.
The big picture here is that last year was about programmers discovering agentic AIs. This year, the business world is following. And there will be a lot of drama of people over doing it, making mistakes, etc. And lots of people whining about how they need to change and insisting that they shouldn't have to. Etc. But this stuff is clearly happening if you look through the noise a bit.
You are absolutely right!
Kidding, but the analogy sits comfortably with me.
I wonder though if this kind of behavior is potentially harmful, most likely less than drugs but nonetheless...
I've been using Siri (via homekit) to turn all my lights on and off for about 3 years now. It's steadily getting worse and worse as somehow, Siri is becoming less accurate and Apple is failing to adopt this new technology in a timely fashion.
I would like to tell it to turn off certain light in a certain room, but unless I get the exact string name of those light correct when I speak, Siri doesn't know what I'm talking about. And it can't do multiple things in a command. I can't say "turn off all the lights in XYZ room" or turn of "this light and this light".
Meanwhile, I can vaguely tell a computer behind my tv to do very complicated things (build me an service that ...) and it can execute on it fairly well. But in apple's "product vision" which I am apparently too dumb to decide for myself what I want, I can't ask for two lights to be turned off at the same time.
How does git replace a vector db search exactly? They are orthogonal. Are you gonna burn a million tokens every time you wanna find some relevant files?
I see the authors don't argue it well enough, but one could use AI agents with simple grep and that proved to be efficient enough to be the default in Claude Code.
I personally turned of indexing feature in Cursor and I use it without it - I haven't noticed any accuracy drop, though my codebase is not enterprise-size one.
Haven't they already proven to be extremely useful? In some areas they are definitely here to stay, coding/software and search (retrieve and summarize information). There's a bunch of places where they are surely shoehorned in, overhyped, and don't belong, but there's also equally many places where they might still be transformative but aren't used yet.
But overall I think the technology is well proven.
reply