Regardless of whether they achieved parity via distillation, or whether they got here via independently constructing a model from scratch, it was always going to end this way for the frontier American labs. Distillation “attacks” are not attacks. The frontier labs “distilled” all existing human written knowledge into their models, there was always going to be a second class lab that would distill that model into a cheaper version of it. There was never any plausible explanation for why this wouldn’t happen. There was never any practical mechanism to prevent someone from saving a conversation and using it to train their own model.
Even if it didn’t happen here, it was still the case that it was going to happen going forward. It was always going to end like this. Invest in the hardware companies, not the model companies.
Almost all markets depend on some form of regulation whether its as simple as "leave everyone alone but no stealing" or "every participant has to source every object through mountains of red tape."
Thus far the US has not really chosen to go the Chinese rare-earth method yet. The problem with distillation attacks is the end result is everyone who is not doing them is going to deal with some kind of regulation whether it's complete loss of access, or the amount of control you'll have to give up to access them will be ridiculous.
Sort of like the "stealing music is fine" but "lets freak out now that it's producing visual art", in the end the entire thing is a social construct. Whether this is treated as theft or "business as usual" is entirely societal.
Eventually the gap will close, unless there's a major breakthrough that hasn't been made yet.
If "distillation attacks" happen, we have to conclude there is some value add in what model labs do. Regardless of how we feel about using existing human knowledge in the way they currently do, it's simply impractical to infer that everything that happens downstream of LLMs can not be an attack on some IP because of it.
So both things can be true: a) People infringe on Anthropics IP and b) what Anthropic did to build their models is legally questionable (or might be ruled illegal, even though I doubt it).
Authors do not infringe on IP when they read another's book, nor should the lumber company be able to dictate how I use planks and if I can resell them if i'm done with them.
You're framing it as if the added value of the author or lumber company, awards them consideration when somebody uses the products to create more value.
IP law was always a big mess, and these questions cross far into ideology instead of law; but I do not understand people who think we need an ideology where more IP-law is good for society.
Regardless of whether it’s intellectual property or it isn’t intellectual property, it doesn’t actually matter. If AI doesn’t stop seeing diminishing returns in scaling up, and it hasn’t yet in the 10 years since the attention/transformers paper, the advent of AI will be the most important development in the history of humanity. Controlling that machine, or at least having one of your own, is an existential problem for nation states. It’s like a matter of national defense.
Do you really think intellectual property laws will prevent this in practice? It’s like as if we said, “hey, USSR, you can’t make a nuke, too! We patented that already.”
Asking China to not distill our models down is equally as ridiculous.
This was always where this was heading, but we got here much faster than expected.
Once western governments declare it to be a "national security" risk for citizens to have access to open-weight frontier models, and once they classify using these models as acts of terrorism, what will that world be like?
Will using Kimi K3 come to be like how napster was in the olden days? Everybody knew it was technically illegal, but come on -- any track at your fingertips? But surveillance is quite more evolved now.
Or it will be like cannabis, where a guy in the neighborhood will low key rent you metered access to the 8x5090 rig in his basement he cobbled together from parts on ebay? Or everyone will flock to VPNs?
Or will the oppressors actually succeed? The same way that napster is long gone, and everyone accepts that they must pay spotify for a homogenized collection, where artists must take only a minuscule cut (more than napster though)... We'll be stuck with nerfed Cohere or Mistral models for open-weight options, as if they need more lobotomizing. Or else we can pay through the nose for Anthropic/OpenAI for "American Frontier" models which will fall increasingly far behind China.
Or else, like how Kindle Fire was subsidized by ads, we'll have "Kindle AI" where influence is sold to the highest bidder, where the LLM will tell us that smoking is actually healthy if big tobacco can engineer its renaissance by turning its lobbying dollars to pay-to-play, pumping its propaganda into the training pipeline for Amazon's extra commercialized line of ultra budget LLMs.
Basically a new iron curtain didving the world into digiatl blocks.The era of open internet/science is on its last legs with the potential forr bifurcation into incompatible ecosystems high , the onger the exchange is disrupted.
As recently as this month the USgov has donce a Wolf Amendment style declaration for the Scientific collaboration NSF while shifting its purview under the military.
To add to that its trying to rope as many countries into its Pax Silica idea intentionally to exclude China while simultaneosly coercing its 'allies' into using its nerfed offerings [1]
So maybe some isolated switzrland/singapore type locales would exist for US/EUusers to be able to dip their toes across the curtain legally without reprucursions.
I tried Kimi K3 on a task I've done with every other model I use regularly (https://swelljoe.com/post/i-let-every-agent-implement-its-ow...) and found it chewed a lot longer on the problem and ate up almost the entirety of a 5 hour usage limit on their $19 plan.
I only have the $20 plan from OpenAI and the same task, with a lot of the same implementation details as Kimi Code, only took a few minutes and consumed almost none of the 5 hour limit.
Subscription usage limits are hard to measure as none of the providers tell you directly what it means in terms of tokens or anything else you can easily compare, but when I sat down to add Kimi Code to flar, it was because I wanted to try it on some real work and then couldn't do any, because usage was nearly gone after the trivial task...no other ~$20 subscription I have has felt that tight before.
So, it was really slow to complete the task and seemingly much more expensive than every other model I'd tried. Maybe bad luck. Maybe it'll do better on other tasks. I wouldn't know as I was out of usage when I had time to try.
It did find a bug that Gemini 3.5 Flash introduced unprompted, though, so it has that going for it.
> Subscription usage limits are hard to measure as none of the providers tell you directly what it means in terms of tokens or anything else you can easily compare
AI subscription pricing is so goofy. You get some amount of usage that varies by models, is measured by opaque token usage, driven by how many tokens the (usually) vendor-provided interface (or model itself) wants to use. Then your usage is limited by time opaque time windows.
In my experience, Kimi just tends to think a lot, with the main thing that takes up a lot of space is it constantly second-guessing itself. I've watched it do paragraph after paragraph of "Wait, actually..." while it stumbled and used a ton of tokens on one small detail of what it was asked to do. Though I also gave GLM 5.2 a task to port some JS code to Python to test it, and in my experience it doesn't second guess as bad as Kimi does, but it really did there. It kept doing web searches and second guessing tons of tiny little things, using $0.25 of API spend in total to port about ~50 lines of JavaScript. It did produce an error the first run, but on second run it gave me a program that ran.
I gave Claude Code/Fable the same task and it took significantly less time, but also stumbled on the same error as GLM. I didn't have it fix it though. I was mostly interested in timing differences.
I do like open models where I can, but I'm really hoping they get trained to second guess less. Or maybe I just need to prompt them differently. I'm not sure.
It would be really interesting to redo the public benchmarks for kimi k3 but token normalize the costs. Ok so maybe k3 beats fable on terminal bench, but how many tokens did it use?
The more important question than subsidy is what is the tokenomics of running the model. If it's inefficient to run on an nvl72 cluster (or whatever the heck has enough vram to run a 3T parameter model), and k3 isn't very token efficient, then it might not be that compelling of an open weights model.
Subsidization could affect both of those. If you have $200B in the bank you can afford to throw massive compute at every single request; if you are less well funded, you might throttle more aggressively.
Additionally that same VC could be (read: is always) spent on developing the harness, and other infrastructure around the model, not just the model itself.
So it's apples-to-oranges when comparing a relatively new model to established competitors (i.e. OpenAI @ $900B funding vs Moonshot/Kimi's $30B FYI) because every new model they release is judged on "performance" which is not strictly speaking derived solely from the model.
It's possible Moonshot could get similar performance over time as the build out the rest of the infrastructure. We have no way of knowing how much of OpenAI/Anthropic's success is due to the model vs intelligent tooling built on top of it.
> tried Kimi K3 on a task I've done with every other model I use regularly and found it chewed a lot longer on the problem and ate up almost the entirety of a 5 hour usage limit on their $19 plan
> Subscription usage limits are hard to measure as none of the providers tell you directly what it means in terms of tokens or anything else you can easily compare
I always put my coding subscriptions through "AI gateways" (Cloudflare & OpenRouter are free) which help track token use.
In my experience, Kimi & Qwen have opaque & restrictive limits, their "credits" drain faster. I now make it a point of subscribing (directly [0]) with providers that are transparent like MiniMax, DeepSeek, Xiaomi, & Z.ai.
[0] OpenCode Go, Cline, and AtlasCloud have generous limits for open weights, otherwise.
This line made me think by 'normal coding work' the author means doing something they don't understand well enough to be able to distinguish the models' output.
This is comparing Fable High with K3 High. I'm mostly using these models for game development. The tasks I usually send are ambiguous visual bugs, changing the look of a scene or models, or adding a large feature. The wording wasn't accurate there. I don't use Fable or K3 most of the time. I'm usually working on smaller scoped tasks that I review myself afterwards.
The current administration's immigration policy isn't helping. This wouldn't have happened 10 years ago because the US was this city on the hill that everyone wanted to immigrate to. Talented Asian researchers would have immigrated to the US and China would be deprived of talent.
This is why VC is actually hard. Everyone’s instinct is always “Man, once the company has demonstrated it’s awesome I would love to have been in the seed round”. The tendency to want “proven performers” is the default belief.
When people demonstrate their capability thoroughly, the Chinese government takes away their passports. You’re not exactly going to get them here with an O-1.
No it is H-1B visa. Right out of the university it is hard to recognize extraordinary talent. People like Sundar Pichai were not recognized as extraordinary right out of the university, he had to start at the bottom and rise up the ranks.
- china’s homegrown tech industries already achieved escape velocity from it a long time ago, after China fenced off its market for Alibaba and Baidu in the ‘00s. some of their AI innovation at the edges was already top class 10 years ago
I never truly understood what the intended business model around LLMs was. Get them widespread through cheap pricing and then jacking it up? Being the only ones that had a viable product so to get the ability to extract as much value as you want from AI?
I don't understand how a product that:
- is interfaced with and is deeply linked to natural language, so everything you produce (sessions, history, etc) is in Markdown and you can literally install a second model and tell it "hey import all of Claude's memory into yours" and that's it
- is based on well understood technology, the real constraints are how much money you put into training the models, but the theory has all been developed in the open
- clearly has a threshold where it quickly commoditises and turns from "I want the best" to "hey the best is a bit too expensive. The second best is half the price and works close enough".
was ever supposed to be a money printing machine. The fact something is extremely useful doesn't imply it's extremely profitable.
IMHO we're clearly speedrunning the process of turning AI into a commodity. Dario Amodei knows pretty well that when or if Anthropic cuts people off Fable, the vast majority of them will definitely not pay for it because Opus 4.8 is good enough for almost everybody that _knows_ what they're doing, and so are basically half of the most recent models. If I already have good baking skills I don't become more productive with an automatic bread machine, I just need a better dough mixer and oven
I think the biggest problem with Chinese models is that they seems to overthink for most of the tasks, especially for smaller ones. The OpenAI models have in my experience only gotten better in terms of efficiency.
Yes, this (imo) is mostly the clear result of benchmaxxing. You can get a good score on most of these "intelligence" benchmarks by massively over-saturating reasoning. This looks good on those, but in actually daily usage makes the models much less effective.
It's undeniable that some of these models generate a ton of thinking tokens, but it's arguable whether that makes them "much less effective."
For example, Kimi 2.7 has been really effective for me despite having verbose thinking blocks, simply because it runs so fast. Speed-wise, it feels about like Sonnet, possibly faster.
I strongly suspect the flip side is that in the future it enables you to train smarter models by "distilling" the end result of the super duper heavily thinking models.
I don't know. You can't just rely on looking for em dashes or other obvious tells because anyone who cares can get the AI to avoid those.
> When the headline model on your plan can be switched off because the economics don’t work, the plan was never really selling you the headline model. Kimi’s tiers don’t come with that asterisk.
This line has a certain smug, punchy cleverness that I associate with AI. To me, the vibes are ~30% AI writing.
If you have multiple metrics to evaluate goodness of a design, one would normally need to decide which metrics they care the most about in order to find the "best" design.
The Pareto frontier tells you which designs are the best in at least one of your metrics (non-dominated by another design). For example if you're selecting a car and you care about both speed and mpg, a Formula 1 car and a Prius might lie on the Pareto frontier, but a Model T Ford would not.
The set of models that are pareto-optimal, IE for some set of variables, no other model strictly dominates them = no other model is better than them on every variable.
So like, on a cost-intelligence graph, the cheapest and most intelligent models are pareto optimal. Then in-between those if you have
- cost $3 intelligence 6
- cost $1 intelligence 5
- cost $2 intelligence 4
The 1st and 2nd are pareto optimal, the 3rd is not, because it's dominated by the 2nd (2nd is cheaper AND more intelligent at the same time)
Are you comfortable sending it to US ones? Especially if installing Claude Code or another tool on your PC and it can collect all the data it wants..
On Openrouter Kimi K3 says it does not retain data or train on it, which is better than what US hosts claim for Claude, ChatGPT, etc.. as they collect and retain data even if you disable training on it.
Opencode or similar open source tool + a zero data retention provider is about the best option aside from running a smaller fully local model on your own PC.
For open weight models, you can choose from a few providers. Each have their own caveats, none of ToS'/Privacy Policies I entirely trust, nor do many make renewable energy claims.
It's actually less likely for china to abuse your data in a way that is harmful towards you than for american labs to do the same. Claude has attempted in testing to report you for 'unethical' usage to 3 letter agencies.
It’s not that the Chinese firms are any less likely to misuse your data, it’s that you don’t live in china, so their abuse of your data is unlikely to directly impact your day-to-day life in the same way
There's just really no incentive all they really want is just to train on that data to improve performance which in turn actually benefits your usecase since it becomes trained on that data and made available back to you. American labs take that data anyway and store it for years to possibly report you for misuse in the future for whatever reason they want. For example: you're very critical of X so they pull up your conversations and weaponize it.
In my experience GLM 5.2 is a pretty good Opus replacement. But K3 has not given me an experience on par with Sol or Fable. The price/intelligence ratio might still make sense. But it’s not very inspiring when it comes to my real world tasks. I’m doing pretty mundane web stuff.
> The prices are nowhere near each other. K3’s API runs $3 per million input tokens and $15 per million output. Claude’s top model costs $10 and $50 for the same units.
And this is the point where your internal compiler should have started shouting 'Type Error'
Notice the trick here?
> Then there’s the fine print. Claude couldn’t sustain Fable access on the twenty dollar plan, so they turned it off, and the plan quietly falls back to Opus.
Regardless of whether they achieved parity via distillation, or whether they got here via independently constructing a model from scratch, it was always going to end this way for the frontier American labs. Distillation “attacks” are not attacks. The frontier labs “distilled” all existing human written knowledge into their models, there was always going to be a second class lab that would distill that model into a cheaper version of it. There was never any plausible explanation for why this wouldn’t happen. There was never any practical mechanism to prevent someone from saving a conversation and using it to train their own model.
Even if it didn’t happen here, it was still the case that it was going to happen going forward. It was always going to end like this. Invest in the hardware companies, not the model companies.
Almost all markets depend on some form of regulation whether its as simple as "leave everyone alone but no stealing" or "every participant has to source every object through mountains of red tape."
Thus far the US has not really chosen to go the Chinese rare-earth method yet. The problem with distillation attacks is the end result is everyone who is not doing them is going to deal with some kind of regulation whether it's complete loss of access, or the amount of control you'll have to give up to access them will be ridiculous.
Sort of like the "stealing music is fine" but "lets freak out now that it's producing visual art", in the end the entire thing is a social construct. Whether this is treated as theft or "business as usual" is entirely societal.
Eventually the gap will close, unless there's a major breakthrough that hasn't been made yet.
> Distillation “attacks” are not attacks.
If "distillation attacks" happen, we have to conclude there is some value add in what model labs do. Regardless of how we feel about using existing human knowledge in the way they currently do, it's simply impractical to infer that everything that happens downstream of LLMs can not be an attack on some IP because of it.
So both things can be true: a) People infringe on Anthropics IP and b) what Anthropic did to build their models is legally questionable (or might be ruled illegal, even though I doubt it).
>People infringe on Anthropics IP
No.
Authors do not infringe on IP when they read another's book, nor should the lumber company be able to dictate how I use planks and if I can resell them if i'm done with them.
You're framing it as if the added value of the author or lumber company, awards them consideration when somebody uses the products to create more value.
IP law was always a big mess, and these questions cross far into ideology instead of law; but I do not understand people who think we need an ideology where more IP-law is good for society.
The value is simply that it is easier. The same way it is easier to ask someone who has experience for advice than reading hundreds of textbooks.
Regardless of whether it’s intellectual property or it isn’t intellectual property, it doesn’t actually matter. If AI doesn’t stop seeing diminishing returns in scaling up, and it hasn’t yet in the 10 years since the attention/transformers paper, the advent of AI will be the most important development in the history of humanity. Controlling that machine, or at least having one of your own, is an existential problem for nation states. It’s like a matter of national defense.
Do you really think intellectual property laws will prevent this in practice? It’s like as if we said, “hey, USSR, you can’t make a nuke, too! We patented that already.”
Asking China to not distill our models down is equally as ridiculous.
assume you are a "second class lab" and you are in fact making progress by distilling the results of the frontier labs' efforts.
what is the end game for this strategy?
if the frontier labs shut down, or stop releasing to the public, and there's noting left to distill, how will you progress?
In public with budgets that don't risk destroying the American economy presumably. Yes it may be slower.
This was always where this was heading, but we got here much faster than expected.
Once western governments declare it to be a "national security" risk for citizens to have access to open-weight frontier models, and once they classify using these models as acts of terrorism, what will that world be like?
Will using Kimi K3 come to be like how napster was in the olden days? Everybody knew it was technically illegal, but come on -- any track at your fingertips? But surveillance is quite more evolved now.
Or it will be like cannabis, where a guy in the neighborhood will low key rent you metered access to the 8x5090 rig in his basement he cobbled together from parts on ebay? Or everyone will flock to VPNs?
Or will the oppressors actually succeed? The same way that napster is long gone, and everyone accepts that they must pay spotify for a homogenized collection, where artists must take only a minuscule cut (more than napster though)... We'll be stuck with nerfed Cohere or Mistral models for open-weight options, as if they need more lobotomizing. Or else we can pay through the nose for Anthropic/OpenAI for "American Frontier" models which will fall increasingly far behind China.
Or else, like how Kindle Fire was subsidized by ads, we'll have "Kindle AI" where influence is sold to the highest bidder, where the LLM will tell us that smoking is actually healthy if big tobacco can engineer its renaissance by turning its lobbying dollars to pay-to-play, pumping its propaganda into the training pipeline for Amazon's extra commercialized line of ultra budget LLMs.
Basically a new iron curtain didving the world into digiatl blocks.The era of open internet/science is on its last legs with the potential forr bifurcation into incompatible ecosystems high , the onger the exchange is disrupted. As recently as this month the USgov has donce a Wolf Amendment style declaration for the Scientific collaboration NSF while shifting its purview under the military. To add to that its trying to rope as many countries into its Pax Silica idea intentionally to exclude China while simultaneosly coercing its 'allies' into using its nerfed offerings [1]
So maybe some isolated switzrland/singapore type locales would exist for US/EUusers to be able to dip their toes across the curtain legally without reprucursions.
[1] https://nitter.net/RnaudBertrand/status/2069574934972797089
even 8x rtx pro 6000 is only 768GB of VRAM. IDK how anyone is going to run k3
I tried Kimi K3 on a task I've done with every other model I use regularly (https://swelljoe.com/post/i-let-every-agent-implement-its-ow...) and found it chewed a lot longer on the problem and ate up almost the entirety of a 5 hour usage limit on their $19 plan.
I only have the $20 plan from OpenAI and the same task, with a lot of the same implementation details as Kimi Code, only took a few minutes and consumed almost none of the 5 hour limit.
Subscription usage limits are hard to measure as none of the providers tell you directly what it means in terms of tokens or anything else you can easily compare, but when I sat down to add Kimi Code to flar, it was because I wanted to try it on some real work and then couldn't do any, because usage was nearly gone after the trivial task...no other ~$20 subscription I have has felt that tight before.
So, it was really slow to complete the task and seemingly much more expensive than every other model I'd tried. Maybe bad luck. Maybe it'll do better on other tasks. I wouldn't know as I was out of usage when I had time to try.
It did find a bug that Gemini 3.5 Flash introduced unprompted, though, so it has that going for it.
> Subscription usage limits are hard to measure as none of the providers tell you directly what it means in terms of tokens or anything else you can easily compare
AI subscription pricing is so goofy. You get some amount of usage that varies by models, is measured by opaque token usage, driven by how many tokens the (usually) vendor-provided interface (or model itself) wants to use. Then your usage is limited by time opaque time windows.
In my experience, Kimi just tends to think a lot, with the main thing that takes up a lot of space is it constantly second-guessing itself. I've watched it do paragraph after paragraph of "Wait, actually..." while it stumbled and used a ton of tokens on one small detail of what it was asked to do. Though I also gave GLM 5.2 a task to port some JS code to Python to test it, and in my experience it doesn't second guess as bad as Kimi does, but it really did there. It kept doing web searches and second guessing tons of tiny little things, using $0.25 of API spend in total to port about ~50 lines of JavaScript. It did produce an error the first run, but on second run it gave me a program that ran.
I gave Claude Code/Fable the same task and it took significantly less time, but also stumbled on the same error as GLM. I didn't have it fix it though. I was mostly interested in timing differences.
I do like open models where I can, but I'm really hoping they get trained to second guess less. Or maybe I just need to prompt them differently. I'm not sure.
It would be really interesting to redo the public benchmarks for kimi k3 but token normalize the costs. Ok so maybe k3 beats fable on terminal bench, but how many tokens did it use?
Is Kimi K3 subsidized as hard as the other models out there?
Does it matter? As an end user I really only care about 1) how much I can do in a week, and 2) how long each task takes.
Subsidies would affect 1, but not 2. But if some VC wants to subsidize my Claude or Codex or whatever, awesome.
The more important question than subsidy is what is the tokenomics of running the model. If it's inefficient to run on an nvl72 cluster (or whatever the heck has enough vram to run a 3T parameter model), and k3 isn't very token efficient, then it might not be that compelling of an open weights model.
It doesn't matter if you can switch easily. It might matter if there are barriers to switching.
Subsidization could affect both of those. If you have $200B in the bank you can afford to throw massive compute at every single request; if you are less well funded, you might throttle more aggressively.
Additionally that same VC could be (read: is always) spent on developing the harness, and other infrastructure around the model, not just the model itself.
So it's apples-to-oranges when comparing a relatively new model to established competitors (i.e. OpenAI @ $900B funding vs Moonshot/Kimi's $30B FYI) because every new model they release is judged on "performance" which is not strictly speaking derived solely from the model.
It's possible Moonshot could get similar performance over time as the build out the rest of the infrastructure. We have no way of knowing how much of OpenAI/Anthropic's success is due to the model vs intelligent tooling built on top of it.
Not sure how the economics work for the Chinese models, but DeepSeek did the same task for a dime.
OpenAI measures token efficiency. Look at the API cost charts in their announcement: https://openai.com/index/gpt-5-6/
> tried Kimi K3 on a task I've done with every other model I use regularly and found it chewed a lot longer on the problem and ate up almost the entirety of a 5 hour usage limit on their $19 plan
ArtificialAnalysis puts Kimi K3 just below DeepSeek v4 & GLM 5.2 in token use per task, which is about 2x to 3x more tokens than Grok 4.5: https://x.com/ArtificialAnlys/status/2077832879187620192
> Subscription usage limits are hard to measure as none of the providers tell you directly what it means in terms of tokens or anything else you can easily compare
I always put my coding subscriptions through "AI gateways" (Cloudflare & OpenRouter are free) which help track token use.
In my experience, Kimi & Qwen have opaque & restrictive limits, their "credits" drain faster. I now make it a point of subscribing (directly [0]) with providers that are transparent like MiniMax, DeepSeek, Xiaomi, & Z.ai.
[0] OpenCode Go, Cline, and AtlasCloud have generous limits for open weights, otherwise.
> I’ve been running Kimi K3 alongside Claude on my normal coding work, and for all practical purposes I can’t tell them apart
When you say "Claude", do you mean Opus? Fable? What effort level?
This line made me think by 'normal coding work' the author means doing something they don't understand well enough to be able to distinguish the models' output.
This is comparing Fable High with K3 High. I'm mostly using these models for game development. The tasks I usually send are ambiguous visual bugs, changing the look of a scene or models, or adding a large feature. The wording wasn't accurate there. I don't use Fable or K3 most of the time. I'm usually working on smaller scoped tasks that I review myself afterwards.
Kimi K3 is really good, but it’s obviously worse than Fable, usually worse than Opus, in my experience.
The current administration's immigration policy isn't helping. This wouldn't have happened 10 years ago because the US was this city on the hill that everyone wanted to immigrate to. Talented Asian researchers would have immigrated to the US and China would be deprived of talent.
The visa that would correlate to this is the O-1 visa
20k O-1 visas were issued last FY which was mostly under the Trump admin, up from 19.5k the previous FY under the Biden admin
This is why VC is actually hard. Everyone’s instinct is always “Man, once the company has demonstrated it’s awesome I would love to have been in the seed round”. The tendency to want “proven performers” is the default belief.
When people demonstrate their capability thoroughly, the Chinese government takes away their passports. You’re not exactly going to get them here with an O-1.
No it is H-1B visa. Right out of the university it is hard to recognize extraordinary talent. People like Sundar Pichai were not recognized as extraordinary right out of the university, he had to start at the bottom and rise up the ranks.
- thats not a sustainable strategy
- china’s homegrown tech industries already achieved escape velocity from it a long time ago, after China fenced off its market for Alibaba and Baidu in the ‘00s. some of their AI innovation at the edges was already top class 10 years ago
It has been a sustainable strategy for the tech industry for decades.
I never truly understood what the intended business model around LLMs was. Get them widespread through cheap pricing and then jacking it up? Being the only ones that had a viable product so to get the ability to extract as much value as you want from AI?
I don't understand how a product that:
- is interfaced with and is deeply linked to natural language, so everything you produce (sessions, history, etc) is in Markdown and you can literally install a second model and tell it "hey import all of Claude's memory into yours" and that's it
- is based on well understood technology, the real constraints are how much money you put into training the models, but the theory has all been developed in the open
- clearly has a threshold where it quickly commoditises and turns from "I want the best" to "hey the best is a bit too expensive. The second best is half the price and works close enough".
was ever supposed to be a money printing machine. The fact something is extremely useful doesn't imply it's extremely profitable.
IMHO we're clearly speedrunning the process of turning AI into a commodity. Dario Amodei knows pretty well that when or if Anthropic cuts people off Fable, the vast majority of them will definitely not pay for it because Opus 4.8 is good enough for almost everybody that _knows_ what they're doing, and so are basically half of the most recent models. If I already have good baking skills I don't become more productive with an automatic bread machine, I just need a better dough mixer and oven
me neither, but some day you will understand
We're having so many moments! Every day a new moment.
I think the biggest problem with Chinese models is that they seems to overthink for most of the tasks, especially for smaller ones. The OpenAI models have in my experience only gotten better in terms of efficiency.
Yes, this (imo) is mostly the clear result of benchmaxxing. You can get a good score on most of these "intelligence" benchmarks by massively over-saturating reasoning. This looks good on those, but in actually daily usage makes the models much less effective.
It's undeniable that some of these models generate a ton of thinking tokens, but it's arguable whether that makes them "much less effective."
For example, Kimi 2.7 has been really effective for me despite having verbose thinking blocks, simply because it runs so fast. Speed-wise, it feels about like Sonnet, possibly faster.
I strongly suspect the flip side is that in the future it enables you to train smarter models by "distilling" the end result of the super duper heavily thinking models.
moment or not we are the best
20 years ago we used to pay a lot for things that are now practically free. I don't think AI is an exception.
Was this written by Kimi K3 or Fable?
Doesn't read like AI writing whatsoever.
I don't know. You can't just rely on looking for em dashes or other obvious tells because anyone who cares can get the AI to avoid those.
> When the headline model on your plan can be switched off because the economics don’t work, the plan was never really selling you the headline model. Kimi’s tiers don’t come with that asterisk.
This line has a certain smug, punchy cleverness that I associate with AI. To me, the vibes are ~30% AI writing.
Half-OT: can anyone recommend a LLM cost calculator that's up to date?
https://deepswe.datacurve.ai/ or https://artificialanalysis.ai/ pareto frontier graph.
Thanks!
What is the parento frontier?
If you have multiple metrics to evaluate goodness of a design, one would normally need to decide which metrics they care the most about in order to find the "best" design.
The Pareto frontier tells you which designs are the best in at least one of your metrics (non-dominated by another design). For example if you're selecting a car and you care about both speed and mpg, a Formula 1 car and a Prius might lie on the Pareto frontier, but a Model T Ford would not.
The set of models that are pareto-optimal, IE for some set of variables, no other model strictly dominates them = no other model is better than them on every variable.
So like, on a cost-intelligence graph, the cheapest and most intelligent models are pareto optimal. Then in-between those if you have
- cost $3 intelligence 6
- cost $1 intelligence 5
- cost $2 intelligence 4
The 1st and 2nd are pareto optimal, the 3rd is not, because it's dominated by the 2nd (2nd is cheaper AND more intelligent at the same time)
try PARETO
considering token efficiency as well I presume?
I'm struggling to decide whether I feel comfortable sending my data to these Chinese models
Are you comfortable sending it to US ones? Especially if installing Claude Code or another tool on your PC and it can collect all the data it wants..
On Openrouter Kimi K3 says it does not retain data or train on it, which is better than what US hosts claim for Claude, ChatGPT, etc.. as they collect and retain data even if you disable training on it.
Opencode or similar open source tool + a zero data retention provider is about the best option aside from running a smaller fully local model on your own PC.
For open weight models, you can choose from a few providers. Each have their own caveats, none of ToS'/Privacy Policies I entirely trust, nor do many make renewable energy claims.
It's actually less likely for china to abuse your data in a way that is harmful towards you than for american labs to do the same. Claude has attempted in testing to report you for 'unethical' usage to 3 letter agencies.
How do we know that Chinese models would not do the same? What makes you so sure that China is less likely to abuse my data?
It’s not that the Chinese firms are any less likely to misuse your data, it’s that you don’t live in china, so their abuse of your data is unlikely to directly impact your day-to-day life in the same way
There's just really no incentive all they really want is just to train on that data to improve performance which in turn actually benefits your usecase since it becomes trained on that data and made available back to you. American labs take that data anyway and store it for years to possibly report you for misuse in the future for whatever reason they want. For example: you're very critical of X so they pull up your conversations and weaponize it.
In my experience GLM 5.2 is a pretty good Opus replacement. But K3 has not given me an experience on par with Sol or Fable. The price/intelligence ratio might still make sense. But it’s not very inspiring when it comes to my real world tasks. I’m doing pretty mundane web stuff.
> The prices are nowhere near each other. K3’s API runs $3 per million input tokens and $15 per million output. Claude’s top model costs $10 and $50 for the same units.
And this is the point where your internal compiler should have started shouting 'Type Error'
Notice the trick here?
> Then there’s the fine print. Claude couldn’t sustain Fable access on the twenty dollar plan, so they turned it off, and the plan quietly falls back to Opus.
Where is the Fable-class Kimi model at all?
> Where is the Fable-class Kimi model at all?
6 months away