I feel like algorithmic/architectural breakthroughs are still the area that will show the most wins. The thing is that insights/breakthroughs of that sort that tend to be highly portable. As Meta showed, you can just pay people 10 million to come tell you what they're doing over there at that other place.
Hasn't this been proven true, many times now? Just look at the difference between ChatGPT 3 and 3.5, for example (which used the same dataset). That, and all the top performing models have large gains from thinking, using the exact same weights.
And, all the new research around self learning architectures has nothing to do with the datasets.
Companies always try to make it seem like data is valuable. Attention is valuable. With attention, you get the data for free. What they monetize is attention. Data is a small part to optimize the sale of ads but attention is the important commodity.
Corruption is the only moat. Oligarchs can buy anything and funnel attention and money into it, creating financial success for shareholders despite poor leadership, zero social responsibility, suboptimal ideas and execution (see: Tesla)
Just commit fraud repeatedly while owning the people who run DoJ, easy peasy, no amount of attention or cash flow can displace that.
Attention is not a moat, it's the thing that's in the castle's treasure room. Without something that makes your service sticky attention may well just walk right out the door.
What if the only moat is domains where it’s hard to judge (non superficial) quality?
Code generation, you don’t see what’s wrong right away, it’s only later in project lifecycle that you pay for it. Writing looks good to skim, is embarrassingly bad once you start reading it.
Some things (slides apparently) you notice right away how crappy they are.
I don’t think it’s just better training data, I think LLMs apply largely the same kind of zeal to different tasks. It’s the places where coherent nonsense ends up being acceptable.
I’m actually a big LLM proponent and see a bright future, but believe a critical assessment of how they work and what they do is important.
Reputation for reliability, stability, or any other desired dimension.
Constant visibility in the news (good, neutral, sometimes even bad!)
A consistent attractive story or narrative around the brand.
A consistent selective story or narrative around the brand. People prefer products designed for "them".
On the dark side: intimidation. Ruthless competition, acquisitions, law suits, reputation for dominance, famously deep pockets.
To keep someone is easier. Tiny things hold onto people: An underlying model that delivers results with less irritation/glitches/hoops. Low to no-configuration installs and operation. Windows that open, and other actions that happen, instantly. Simple attention to good design can create fierce loyalty, for those for whom design or friction downgrades feel like torture.
What's annoying is that companies capture user data and then lock it into their platforms, transform it, and resell it. But it is really the user's data that they're selling back to us. I would like regulation here, you capture my data then I can pick who you must and must not share it with.
Data has historically been a moat, but I think now more than ever it's a moat of bounded size / utility.
The biggest data hoarders now compress their data into oracles whose job is to say whatever to whoever - leaking an ever-improving approximation of the data back out.
DeepSeek was a big early example of adversarial distillation, but it seems inevitable to me that frontier models can and will always be siphoned off in order to produce reasonably strong fast-follow grey market competition.
Why is it that we have agents that can prospect for sales leads and answer support tickets accurately, but we don’t seem to be able to consistently generate high quality slides?
I don't know about prospecting, but "answer support tickets accurately"? Seriously, this must be ironic, right?
Efficiency will ultimately decide if LLMs become feasible long-term. Right now, the LLM industry is not sustainable. Investors were promised literally the future in the present and it is now undeniable that ASI, AGI or even moderately competent general purpose quasi-autonomous systems won't happen anytime soon. The reality is that there is not space for all these players in the market in the long-term. LLMs won't go away but the vast majority of mainstream providers will definitely do
Marketing/relationships is the only moat, not data. You can have amazing data and make an amazing product, and some asshat with a product that barely works and really tight marketing will crush you. Then people will ask why there isn't a product like yours on the market, all while ignoring all your marketing material.
You get some anecdotal evidence and immediately post a hot take claiming to have discovered a new invariant?
I guess a bunch of us, including myself have taken the engagement bait here but does it really take somebody saying something stupid to start a conversation on something?
I find the premise that coding is one of the hardest problem for LLMs flawed. Isn't coding the easiest area for AI, with lots of data to train and easily verifiable?
22 comments
[ 2.6 ms ] story [ 42.0 ms ] threadinb4 "then why do Meta's models still suck?"
And, all the new research around self learning architectures has nothing to do with the datasets.
Companies always try to make it seem like data is valuable. Attention is valuable. With attention, you get the data for free. What they monetize is attention. Data is a small part to optimize the sale of ads but attention is the important commodity.
Why else are celebrities so well paid?
Just commit fraud repeatedly while owning the people who run DoJ, easy peasy, no amount of attention or cash flow can displace that.
Also as foundation models improve, today's "hard to solve" problems become tomorrow's "easy to solve" problems
- Which brands do people trust? - Which people do people of power trust?
You can have all the information in the world but if no one listens to you then it’s worthless.
Code generation, you don’t see what’s wrong right away, it’s only later in project lifecycle that you pay for it. Writing looks good to skim, is embarrassingly bad once you start reading it.
Some things (slides apparently) you notice right away how crappy they are.
I don’t think it’s just better training data, I think LLMs apply largely the same kind of zeal to different tasks. It’s the places where coherent nonsense ends up being acceptable.
I’m actually a big LLM proponent and see a bright future, but believe a critical assessment of how they work and what they do is important.
Vertical integration.
Horizontal integration.
Cross- and/or mass-relationship integration.
Individual relationship investment/artifacts.
Reputation for reliability, stability, or any other desired dimension.
Constant visibility in the news (good, neutral, sometimes even bad!)
A consistent attractive story or narrative around the brand.
A consistent selective story or narrative around the brand. People prefer products designed for "them".
On the dark side: intimidation. Ruthless competition, acquisitions, law suits, reputation for dominance, famously deep pockets.
To keep someone is easier. Tiny things hold onto people: An underlying model that delivers results with less irritation/glitches/hoops. Low to no-configuration installs and operation. Windows that open, and other actions that happen, instantly. Simple attention to good design can create fierce loyalty, for those for whom design or friction downgrades feel like torture.
Obviously, many more moats in the physical world.
The biggest data hoarders now compress their data into oracles whose job is to say whatever to whoever - leaking an ever-improving approximation of the data back out.
DeepSeek was a big early example of adversarial distillation, but it seems inevitable to me that frontier models can and will always be siphoned off in order to produce reasonably strong fast-follow grey market competition.
I don't know about prospecting, but "answer support tickets accurately"? Seriously, this must be ironic, right?
You get some anecdotal evidence and immediately post a hot take claiming to have discovered a new invariant?
I guess a bunch of us, including myself have taken the engagement bait here but does it really take somebody saying something stupid to start a conversation on something?
The law even demands that the data is machine readable.
The only real moat is your own, observational data.