86 Comments

This LLM situation is the man crawling around under a streetlight looking for his keys.

Why are you looking here? Because this is where the light is.

Expand full comment

That’s where the Street-LLMp is?

Expand full comment

😜🫵👍

Expand full comment

Ancient Yiddish joke, but well suited to the context.

Expand full comment

I’m grateful to read this oppositional article esp as the UK Government appear ready to throw their paltry contribution against the wall to see if it will stick when in their press briefing they referred to chess engines as the next big thing conjuring an image of lambs to the slaughter of US big lobby.

Expand full comment

This was indeed predictable but the scaling narrative has taken up a lot of oxygen. Combined with the efficiency developments now commercially available (Deepseek. Qwen), they have downward price pressures that do not bode well. Unfortunately though, the scaling narrative keeps asking to double down on the approach potentially at the cost of what really makes AI reliable. It’s about time we ask hard questions about US innovation ecosystem:

https://open.substack.com/pub/mohakshah/p/the-deepseek-saga-are-we-taking-the?r=224koz&utm_medium=ios

Expand full comment

I am taking bets on who's going to be the first pets.com and @home of the AI-bubble era...

Expand full comment

Well, as Gary has predicted on whether OpenAI may end up being the wework of AI, we may see some contenders for such slots. As I mentioned in the Substack article above, it’s amusing also the time when SoftBank decides to invest in OpenAI efforts. So, history may seem to start rhyming

Expand full comment

Just subscribed to your stack - thanks for sharing.

Expand full comment

Thanks. Hope you’ll find it useful.

Expand full comment

Initial read further confirms my thoughts on AI and HealthTech (in US context): what a waste of money, none of this is going to help un-insured get insured. None of it.

Expand full comment

Indeed. We are succumbing to rhetoric. What we need is to care about the outcomes. You may find this useful:

https://open.substack.com/pub/mohakshah/p/ai-policy-we-need-to-urgently-pivot?r=224koz&utm_medium=ios

Expand full comment

The Humane AI pin is a pretty strong contender for being the thing that most exemplifies the AI bubble. A nearly useless product that cost far too much and didn't do anything your smart phone didn't do already, but was subject to crazy amounts of hype.

Expand full comment

Yes! it is indeed! When it came out 9 months ago, my mind went, oh yeah dot-com all over again

Expand full comment

The DeepSeek developments are very good news for OpenAI and the rest. Their algorithms have been very rough, and the fact that they took so much compute did not bode well.

That optimizations would be found was fully expected, and is the usual way things operate.

Now the companies can double-down on fidelity. I don't think LLM alone are enough, but it is our best bet at conquering the messy world at a high level. Refinements can then be done with other approaches.

Expand full comment

True but with some caveats. Fidelity is non-trivial due to the inherent LLM nature. So there is likely to be convergence in capability sooner or later (unless there is departure and/or significant improvement in the core paradigm).

So what it comes down to would be cost-capability trade-off’s and that would naturally put a downward pressure on price esp when the capability differences are not huge. If the business model had relied on significant input costs, that model is now under pressure give the sunk costs.

Expand full comment

Altman had the nerve to tell high schoolers more/less that there's no point to "outrunning AI tools" and his advice was to start using the tools and basically told students to ask themselves to use AI to help them do things "faster." Translation: subscribe to our overhyped, overpriced products, give us your data and money and don't bother learning beyond prompting. Look these tools are quite powerful, but they are also very heavily flawed with many limitations that won't readily be "solved" and in many cases not foreseeably, at least not with current LLM frameworks. These tech bros just wanna make society reliant on their substandard software and control us with their platforms and data. Welcome to the new oligarchy.

Expand full comment

OpenAI's search for AGI resembles stumbling around in the dark looking for the light switch.

Expand full comment

Stumbling around in the dark looking for the house?

Expand full comment

Worry not, Inshallah, Elon's Theil-Planatir-Sacks Bro-sphere will tear down this wall, quite literally. j/k but maybe?

Expand full comment

“If we steal all the federal government’s data, we can create AGI, or at least a terrifying surveillance state that can force frightened citizens into pretending to believe we have created AGI.”

Expand full comment

AGI - Artificial General IDIOT. A useful idiot that is still highly flawed, but useful enough to pretend to be smart used to replace or control bigger idiots that buy into tech bros bullshit. Too much of that on the internet in places like YouTube and Reddit already blindly relying on these tools including at work. Now imagine what happens when whole infrastructures rely on these flimsy frameworks. It'll be a worse disaster than American airport infrastructure.

Expand full comment

We've been living in a surveillance state since shortly after September 2001 when Bush Jr signed the so-called Patriot Act. Even under Obama it worsened. Nothing new.

Expand full comment

Not new, but it can get worse, my dude. (Librarians like me have been pointing this out since the Patriot Act).

Expand full comment

Agreed, I just don't see how musk and friends rifling through federal govt records is the thing that makes it worse. That's the govt being surveilled, not the govt surveiling us.

Expand full comment

Have you ever sat through a training on handling personally identifiable information. The government has maybe the most of it.

Expand full comment

Yes, I know all about PII. I work at a place that process terabytes of medical claims data. I'm saying there's no net new data being gathered by the got here, just a different set of people looking at it. I think maybe you're just less fond of the collection of people you imagine have access to it now (do any of us really know who has/had access? Any reason to trust them more or less?). The solution then isn't just wait until someone you like better has access, because the tides will just turn against you again later in our seesaw political system. The solution is for the spooks to stop collecting so damn much data on us. Personally, I think that with someone like Tulsi Gabbard in the highest intelligence position in the country, that's actually *more* likely to happen under this admin, not less. This is the woman who on the floor of Congress refused to call Snowden a traitor despite relentless badgering from senators, and then more than half of those senators voted to confirm her. I'd say that sounds like the tide in Washington is turning towards "let's maybe stop warrantlessly dragnetting the whole country now". I'll take it no matter what letter is next to the person's name who advocates for it.

Expand full comment

Decades ago, expert systems were all the rage. Then they hit a wall because complexity tends to grow exponentially. History is busy repeating itself.

Expand full comment

Indeed. There will be winter and then there will be summer again in ~2045 because the computational efficiency has increased by some two orders of magnitude.

Expand full comment

I still follow Gary even though the message doesn't change much. If you're going to kill a varmint like OpenAI, you can't stop beating it till it stops moving. I appreciate the beating Gary gives OpenAI and its simulacra.

I will look forward to Gary either defining AGI or riding into the sunset with everybody waving and wishing him safe travels. He teases AGI but no one in AI ever says what they would do with one if it was instantiated.

There is no intelligence without goals. "Intelligence" without goals is auto complete or blabbering.

AIs can look like they have goals by the words emitted.

Does anyone believe they or some future AI will have a life force that moves it?

Gary, how about beating on the concept of AGI to finish killing the beast?

Expand full comment

I personally hope AGI is an insurmountable challenge. To make a computer think like a human when we barely understand how humans do their thinking seems impossible to me.

My hope is that when the AI bubble collapses there will be another decades long AI winter of no funding for or interest in true AI research.

Expand full comment

I think it really is a long way off, but ultimately feasible. The assertion is that humans are physical entities entirely bound by the laws of physics, hence: Nothing but biological machines. If that were true, then there is no aprioric limit to what a machine can achieve, including conciousness and intelligence.

However you are completely correct that we have not the slightest idea how the human intellect works. Now yeah, we know about neurons and activations from EEG, MRT and stuff, but there is no scientifically accepted *and* consistent model that would be able to explain what is going on.

So let me end with this:

“If the human brain were so simple that we could understand it, we would be so simple that we couldn’t.”

― Emerson M. Pugh

Expand full comment

The fact that we can't even understand how primitive LLM AI behaves should be a warning for building true AGI. We are using LLMs for all types of nefarious tasks, scams, deceptions etc. AGI is another level of pandora's box.

And in defiance of everyone, I don't perceive alignment is possible at all, ever.

https://www.mindprison.cc/p/ai-alignment-why-solving-it-is-impossible

Expand full comment

A simple mantra: If you convert data into probabilistic word pairs, you irreversibly lose accurate recovery of that data.

Expand full comment

Building AGI was never going to be as easy as what Dario Amodei used to claim. As in having a 100 billion cluster and scaling laws.

Skeptics should not mistake this for lack of progress. It is just going to take time.

The AI offerings improved remarkably in 2024. The "parrot" gained the ability to iterate and reflect on its own work. That was a bigger deal than the GPT-3 to GPT-4 jump.

Further improvements will require painstaking work, including in modeling.

"Neurosymbolic" cannot be the whole solution, as neither neural nets nor symbols have easily defined meaning. So will need honest models, whether for physics, biology, etc. Hard work.

Expand full comment

Looks like De'Nial has become the largest river in California. I hope it doesn't wash away all the electric power in Northern CA.

Expand full comment

Have a think about this. I don't know the cost so far, but surely it would be cheaper for the AI companies to have opened call centres staffed with humans. They could then answer user queries, by manually querying an agi system and retyping the answer with their added human input. Bingo, success. It would also provide employment for the millions of people who lost their jobs due to automation.

Expand full comment

Couldn’t agree more. It seems like some people in AI believes access to unlimited capital and computing power is the key to technology advancement, but in fact innovation always happens because of the need to work within some constraint. Nvidia wouldn’t do chip design in simulation if they were not constrained by time, Google wouldn’t have invented MapReduce and GFS if they weren’t constrained to use cheap hardware. We wouldn’t even have ikea if Ingvar kamprad didn’t insist all designs be constrained by cost. Unlimited access resource might not be the advantage OpenAI deems it to be.

Expand full comment

Clever comment. DeepSeek definitely showed that with limited resources and able to produce such an impressive model.

Expand full comment

The only possible route to actual progress would be just not build NNs and derivative techs such as LLMs i.e. START THE HECK OVER https://davidhsing.substack.com/p/what-the-world-needs-isnt-artificial

Expand full comment

AI companies are trying to “solve” Homer Simpson’s list of ill- and un-defined problems

Expand full comment

People who on know how language and intelligence work seem to have insight that the casual observers and the technology specialists do not in this game.

Expand full comment

A trend is a trend until the bend in the end.

Expand full comment

The proof of the pudding, dear Gary, lies in the eating, and there already is plenty of very nourishing pudding to go around, ask people who have whetted their appetites on deep research and have claimed that the $200 subscription is a throwaway compared to what it achieves. You're betting on the wrong horse here, or rather not betting on anything at all but simply badmouthing the owners of horses that can already run pretty fast. Also you are making the mistake of betting against c, the real workhorse in all this.

Expand full comment

I can't use a product that invents ~10% of the references in its "research summaries".

Expand full comment

Agreed. A chair that doesn’t collapse 80% of the time is not particularly useful.

Expand full comment

That's how I look at it.

Or an accounts receivable employee who correctly bills 87% of my firm's invoicing.

Expand full comment

What about a folding chair?

Expand full comment

Or a folding AI company?

Expand full comment

No one is saying that it is perfect, given finiteness of resources, you can't possibly train a model that achieves absolute perfection, but that's what humans are for, in my opinion the centaur like feature of humans and AI collaborating together is going to continue for quite some time, at least until the time where we have more significant algorithmic breakthroughs that spawn machines that can operate completely autonomously.

Expand full comment

I have nothing against collaboration. I do object to hype.

I occasionally ask the various AI interfaces to solve very simple logical questions. Questions that any normal 11 year old will figure out, perhaps with a little help.

They fail consistently. They give wrong answers with complete confidence and explain how they "solved" the problem. One can "talk them out of" their errors, but why bother?

Expand full comment

Asking transformers to answer questions that it has not encountered in its training data is like trying to judge the merit of an elephant by asking it to climb a tree. Same applies to ARC-AGI type puzzles. Their utility lies elsewhere. Never judge anyone or anything by its failings. Also, it would need further algorithmic breakthroughs.

Expand full comment

I see. Personally, I would expect anything defined as "intelligent" to be able to deal with unaccustomed questions.

LLMs are built on the statistical juxtaposition of words in very large language samples. But, sadly for LLMs, languages are not lexicostatistics.

I just asked the latest version of Google's Gemini this question:

"Alice and Sarah are first cousins. Alice has six first cousins. How many first cousins does Sarah have?"

With the same discouraging results as on previous attempts. Gemini showed, in painful detail, how it arrived at its completely wrong answer.

Expand full comment

"languages are not lexicostatistics"

Linguists have been pointing that out for decades. AI Numerologists have ignored them.

Expand full comment

I would guess as much

Expand full comment

Where, pray tell, does their utility lie?

Expand full comment

AI that can only work with what it has already encountered has a pre existing name: search engine.

So you are saying gpt is a glorified search engine. I agree. Let’s drop the hype then.

Expand full comment

Your anecdotally anonymous folks you know are neither a return on investment nor a studied mass. So we can always find someone who looks good in stripes and plaids but it might not be best for the market or society. And the uptake is not looking good for investors. https://www.runtime.news/big-business-is-still-skeptical-about-ai-agents/?ref=runtime-newsletter

Expand full comment