RAG and LLMs are not the same thing, but 'Agents' incorporate both.
Maybe we could resolve the bit of a conundrum by the op in requiring 'agents' to give credit for things if they did rag them or pull them off the web?
It still doesn't resolve the 'inherent learning' problem.
It's reasonable to suggest that if 'one person did it, we should give credit' - at least in some cases, and also reasonable that if 1K people have done similar things ad the AI learns from that, well, I don't think credit is something that should apply.
But a couple of considerations:
- It may not be that common for an LLM to 'see one thing one time' and then have such an accurate assessment of the solution. It helps, but LLMs tend not to 'learn' things that way.
- Some people might consider this the OSS dream - any code that's public is public and it's in the public domain. We don't need to 'give credit' to someone because they solved something relatively arbitrary - or - if they are concerned with that, then we can have a separate mechanism for that, aka they can put it on Github or Wikipedia even, and then we can worry about 'who thought of it first' as a separate consideration. But in terms of Engineering application, that would be a bit of a detractor.
> if 1K people have done similar things ad the AI learns from that, well, I don't think credit is something that should apply.
I think it should.
Sure, if you make a small amount of money and divide it among the 1000 people who deserve credit due to their work being used to create ("train") the model, it might be too small to bother.
But if actual AGI is achieved, then it has nearly infinite value. If said AGI is built on top of the work of the 1000 people, then almost infinity divided by 1000 is still a lot of money.
Of course, the real numbers are way larger, LLMs were trained on the work of at least 100M but perhaps over a billion of people. But the value they provide over a long enough timespan is also claimed to be astronomical (evidenced by the valuations of those companies). It's not just their employees who deserve a cut but everyone whose work was used to train them.
> Some people might consider this the OSS dream
I see the opposite. Code that was public but protected by copyleft can now be reused in private/proprietary software. All you need to do it push it through enough matmuls and some nonlinearities.
- I don't think it's even reasonable to suggest that 1000 people all coming up with variations of some arbitrary bit of code either deserve credit - or certainly 'financial remuneration' because they wrote some arbitrary piece of code.
That scenario is already today very well accepted legally and morally etc as public domain.
- Copyleft is not OSS, it's a tiny variation of it, which is both highly ideological and impractical. Less than 2% of OSS projects are copyleft. It's a legit perspective obviously, but it hasn't bee representative for 20 years.
Whatever we do with AI, we already have a basic understanding of public domain, at least we can start from there.
"it would struggle to honor its long-term agreements. That failure would cascade. Oracle, for example, could be left with massive liabilities and no matching revenue stream,"
No, there's a not of noise about this but these are just 'statements of intent'.
Oracle very intimately understands OpenAI's ability to pay.
They're not banking $50B in chips and then waking up naively one morning to find out OpenAI has no funding.
What will 'cascade' is maybe some sentiment, or analysts expectations etc.
Some of it, yes, will be a problem - but at this point, the data centre buildout is not an OpenAI driven bet - it's a horizontal be across tech.
There's not that much risk in OpenAI not raising enough to expand as much as it wants.
Frankly - a CAPEX slowdown will hit US GDP growth and freak people out more than anything.
OpenAI is still de facto the market leader in terms of selling tokens.
"zero moat" - it's a big enough moat that only maybe four companies in the world have that level of capability, they have the strongest global brand awareness and direct user base, they have some tooling and integrations which are relatively unique etc..
'Cloud' is a bigger business than AI at least today, and what is 'AWS moat'? When AWS started out, they had 0 reach into Enterprise while Google and AWS had infinity capital and integration with business and they still lost.
There's a lot of talk of this tech as though it's a commodity, it really isn't.
The evidence is in the context of the article aka this is an extraordinary expensive market to compete in. Their lack of deep pockets may be the problem, less so than everything else.
This should be an existential concern for AI market as a whole, much like Oil companies before highway project buildout as the only entities able to afford to build toll roads. Did we want Exxon owning all of the Highways 'because free market'?
Even more than Chips, the costs are energy and other issues, for which Chinese government has a national strategy which is absolutely already impacting the AI market. If they're able to build out 10x data centres at offer 1/10th the price at least for all the non-Frontier LLM, and some right at the Frontier, well, that would be bad in the geopolitical sense.
The AWS moat is a web of bespoke product lock-in and exorbitant egress fees. Switching cloud providers can be a huge hassle if you didn't architect your whole system to be as vendor-agnostic as possible.
If OpenAI eliminated their free tier today, how many customers would actually stick around instead is going to Google's free AI? It's way easier to swap out a model. I use multiple models every day until the free frontier tokens run out, then I switch.
That said, idk why Claude seems to be the only one that does decent agents, but that's not exactly a moat; it's just product superiority. Google and OAI offer the same exact product (albeit at a slightly lower level of quality) and switching is effortless.
There are quite large 'switching costs' from moving a solution that's dependent on on model and ecosystem, to another.
Models have to significantly outperform on some metric in order to even justify looking at it.
Even for smaller 'entrenchements' like individual developers - Gemeni 3 had our attention for all of 7 days, now that Opus 4.5 is out, well, none of my colleagues are talking abut G3 anymore. I mean, it's a great model, but not 'good enough' yet.
I use that as an example to illustrate broader dynamics.
Open AI, Anthropic and Google are the primary participants here, with Grok possibly playing a role, and of course all of the Chinese models being an unknown quantity because they're exceptional in different ways.
Switching a complex cloud deployment from AWS to GCP might take a dedicated team of engineers several months. Switching between models can be done by a single person in an afternoon (often just 5 minutes). That's what we're talking about.
That means that none of these products can ever have a high profit margin. They have to keep margins razor thin at best (deeply negative at present) to stay relevant. In order to achieve the kinds of margins that real moats provide, these labs need major research breakthroughs. And we haven't had any of those since Attention is All You Need.
" Switching between models can be done by a single person in an afternoon (often just 5 minutes). That's what we're talking about."
Good gosh, no, for comprehensive systems it's considerably more complicated than that. There's a lot of bespoke tuning, caching works completely differently etc..
"That means that none of these products can ever have a high profit margin."
No, it doesn't. Most cloud providers operate on a 'basis' of commodity (linux, storage, networking) with proprietary elements, similar to LLMs.
There doesn't need to be any 'breakthroughs' to find broad use cases.
The issue right now is the enormous underlying cost of training and inference - that's the qualifying characteristic that makes this landscape different.
Aren't you contradicting yourself? To even be considering all the various models, the switching cost can't be that large.
I think the issue here isn't really that it's "hard to switch" it's that it's easier yet to wait 1 more week to see what your current provider is cooking up.
But if any of them start lagging for a few months I'm sure a lot of folks will jump ship.
Selling tokens at a massive loss, burning billions a quarter isn't the win you think it is. They don't have a moat bc they literally just lost the lead, you only can have a moat when you are the dominant market leader which they never were in the first place.
> All indications are that selling tokens is a profitable activity for all of the AI companies - at least in terms of compute.
We actually don't this yet because the useful life of the capital assets (mainly NVIDIA GPUs) isn't really well understood yet. This is being hotly debated by wall st analysts for this exact reason.
Gemeni does not have 'the lead' in anything but a benchmark.
The most applicable benchmarks right now are in software, and devs will not switch from Claude Code or Codex to Antigravity, it's not even a complete product.
This again highlights quite well the arbitrary nature of supposed 'leads' and what that actually means in terms of product penetration.
And it's not easy to 'copy' these models or integrations.
I think you're measuring the moat of developing the first LLMs but the moat to care about is what it'll take to clone the final profit generating product. Sometimes the OG tech leader is also the long term winner, many times they are not. Until you know what the actual giant profit generator is (e.g. for Google it was ads) then it's not really possible to say how much of a moat will be kept around it. Right now, the giant profit generator is not seeming to be the number of tokens generated itself - that is really coming at a massive loss.
I mean, on your Cloud point I think AWS' moat might arguably be a set of deep integrations between services, and friendly API's that allow developers to quickly integrate and iterate.
If AWS' was still just EC2, and S3 then I would argue they had very little moat indeed.
Now, when it comes to Generative AI models, we will need to see where the dust settles. But open-weight alternatives have shown that you can get a decent level of performance on consumer grade hardware.
Training AI is absolutely a task that needs deep pockets, and heavy scale. If we settle into a world where improvements are iterative, the tooling is largely interoperable... Then OpenAI are going to have to start finding ways of making money that are not providing API access to a model. They will have to build a moat. And that moat may well be a deep set of integrations, and an ecosystem that makes moving away hard, as it arguably is with the cloud.
EC2 and S3 moat comes from extreme economies of scale. Only Google and Microsoft can compete. You would never be able to achieve S3 profitability because you are not going to get same hardware deals, same peering agreements, same data center optimization advantages. On top of that there is extremely optimized software stack (S3 runs at ~98% utilization, capacity deployed just couple weeks in advance, i.e. if they don’t install new storage, they will run out of capacity in a month).
I wouldn't call it a moat. A moat is more about switching costs rather than quality differentiation. You have a moat when your customers don't want to switch to a competitor despite that competitor having a superior product at a better price.
Why 'host' just to tap a few prompts in and see what happens? Worst case, you loose an account. Usually the answer has to do with people being less sophisticated than otherwise.
Nobody has access to 'frontier quality models' except Open AI, Anthropic, Google, maybe Grok, maybe Meta etc. aka nobody in China quite yet. And - there are 'layers' of Engineering beyond just model that make quite a big difference. For certain tasks, GPT5 might be beyond all others, same for Claude + Claude.
That said, the fact that they're doing this while knowing that Anthropic could be monitoring implies a degree of either real or arbitrary irreverence: either they were lazy or dumb (unlikely), or it was some ad hoc situation wherein they really just did not care. Some sub-sub-sub team at some entity just 'started doing stuff' without a whole lot of thought.
'State Backed Entities' are very numerous, it's not unreasonable that some of them, somewhere are prompting a few things that are sketchy.
I'm sure there's a lot of this going on everywhere - and this is the one Anthropic chose to highlight for whatever reasons, which could be complicated.
> Nobody has access to 'frontier quality models' except Open AI, Anthropic, Google, maybe Grok, maybe Meta etc. aka nobody in China quite yet.
welcome to 2025. Meta doesn't have anything on par with what Chinese got, that is common knowledge. Kimi, GLM, QWen and MiniMax are all frontier models no matter how you judge it. DeepSeek is obviously cooking something big, you need to be totally blind to ignore that.
America's lead in LLM is just weeks, not quarters or years. Arguing that Chinese spy agencies have to rely on American coding agents to do its job is more like a joke.
according to the SWE bench results I am looking at, KIMI K2 has higher agentic coding score than Gemini and its gap with Claude Haiku 4.5 is just 71.3% vs 73.3%, that 2% difference is actually less than the 3% gap between GPT 5.1 (76.3%) vs Claude Haiku 4.5. interestingly, Gemini and Claude Haiku 4.5 are "frontier" according to you but KIMI K2, which actually has the higest HLE nd Live Codebench results, is just "near" the frontier.
You started by saying 'There's no way to judge!' - but then bring out 'Benchmarks!' ... and hypocritically infer that I have 'dual standards'?
The snark and ad hominem really undermine your case.
I won't descend to the level of calling other people names, or their arguments 'A Joke', or use 'It's Common Sense!' as a rhetorical device ...
But I will say that it's unreasonable to imply that Kimi, Qwen etc are 'Frontier Models'.
They are pretty good, and narrowly achieve some good scores on some benchmarks - but they're not broadly consistent at that Tier 1 quality.
They don't have the extended fine tuning which makes them better for many applications, especially coding, nor do they have the extended, non-LLM architecture components that further elevate their usefulness.
Nobody would choose Qwen for coding if they could have Sonnet at the same price and terms.
We use Qwen sometimes because it's 'cheap and good' not because it's 'great'.
The 'true coding benchmark' is that developers would chose Sonnet over Qwen, 99 out of 100 times, which is the difference between 'Tier 1' and 'Not Really Tier 1.
Finally, I run benchmarks with my team and I see in a pretty granular way what's going on.
What I've said above lines up with reality of our benchmarks.
We're looking at deploying with GLM/Z.ai - but not because it's the best model.
Google, OAI and Anthropic score consistently better - the issue is 'cost' and the fact that we can overcome the limitations of GLM. So 'it's good enough'.
That 'real world business case' best characterizes the overall situation.
Most people lean on tradition for ideals. They do what they always have done and what they see people around them do. But if you break new ground as technology does, then that is not possible. You have to use reason and philosophy, and people will come to different conclusions. Those who end up at a non-mainstream conclusion are then labeled culty.
Trump's personal, newfound multi-billion dollar crypto fortune is hosted by Zhao.
I don't mean to be breaking any etiquette her by re-indicating this, but it's I think it's unreasonable to suggest that Trump could not know who this person is.
This is Trump's new 'personal banker' , who doesn't have to play be the constrained rules of $USD denominated financial regulations.
To the contrary, this is possibly an extremely dangerous and corrupt action (I'm saying 'possible' here).
Zhao manages Binance upon which Trump's nefarious crypto operation - World Liberty Financial is hosted.
Zhao is guilty of wilfully ignoring a lot of very bad and illegal activity by all sorts of bad actors.
World Liberty Financial is receiving $100's of millions inbound, coincident with 'deals' made by US and Middle East actors etc..
WLF is co-managed by Steve Witkoff, the 'real estate magnate' who has been charged with Middle Easter nand Russia negotiations (who by the ay has absolutely no diplomatic background, historical context or understanding of this situations and is deeply unqualified) and who notably has been entering into negotiations and discussion with foreign parties without any US State Dept personnel. Sometimes not even translators.
Subsequent to those 'deals' with Qatar etc. which the Administration indicated there would be up to $1T invested in the US ... Qatar and other regimes have been flushing massive amounts of noney into WLF, hosted on Binance, overseen by Zhao.
The potentiality fore corruption is hard to overstate.
> World Liberty Financial is receiving $100's of millions inbound, coincident with 'deals' made by US and Middle East actors etc..
The thing is, in the corruption surrounding the President, that's still small fish. $TRUMP alone was worth 13 billion dollars, $MELANIA was at 1.7 billion dollars. And that's just these two meme coins, not going into all the other shenanigans - he and his family are expected to have made 3 billion dollars since the election [1] in personal wealth gain.
I'm not sure we can differentiate like that. If WLF and Witkoff are corrupt, that means all the big "peace talks" are crypto shakedowns behind closed doors, and suddenly we are back at "Geopolitical Level Corrupt".
Maybe we could resolve the bit of a conundrum by the op in requiring 'agents' to give credit for things if they did rag them or pull them off the web?
It still doesn't resolve the 'inherent learning' problem.
It's reasonable to suggest that if 'one person did it, we should give credit' - at least in some cases, and also reasonable that if 1K people have done similar things ad the AI learns from that, well, I don't think credit is something that should apply.
But a couple of considerations:
- It may not be that common for an LLM to 'see one thing one time' and then have such an accurate assessment of the solution. It helps, but LLMs tend not to 'learn' things that way.
- Some people might consider this the OSS dream - any code that's public is public and it's in the public domain. We don't need to 'give credit' to someone because they solved something relatively arbitrary - or - if they are concerned with that, then we can have a separate mechanism for that, aka they can put it on Github or Wikipedia even, and then we can worry about 'who thought of it first' as a separate consideration. But in terms of Engineering application, that would be a bit of a detractor.
reply