Dankrad Feist

Most definitions boil down to being able to accomplish what a human can accomplish, which makes the above statement kind of trivial. My personal view is that there is probably no such thing as General Intelligence. It's just language. When a human figures something out by chance they are able to tell all the others how it works, so we end up being way smarter than animals.

@dankrad are you saying everyone has a different definition? I would have thought that there are some common definition people agree on

When you use it for a while on a project, it is clear that even Fable misses things that are obvious to you and you need to drive it in the right direction. Not clear to me what it is, but this has so far I have always felt this with new models after the first impression wears off. Don't get me wrong, it already has superhuman performance for a lot of tasks, but it still feels like there is something missing, although it is hard to pinpoint what exactly. Something happens in the human brain as you get deep into a project and get a good understanding, that models just don't as far as I can observe. Will be interesting if (and when) that gap closes.

@cryptodavidw Can you define AGI?

@dankrad Its the AGI that’s missing. We’re in between auto complete and AGI and we’re not used to it. It’s uncanny

@matthewdif See a physio instead

>having pain in my shoulder recently after gym >go to doctor to get checked out >doctor's advice is "lol easiest fix is just stop working out" fml, every time

@matthiasgeihs Maybe. Something used to believe strongly but became less sure of as models advanced. But it could be that online learning is the missing ingredient

> Something happens in the human brain as you get deep into a project and get a good understanding, that models just don't as far as I can observe. Maybe that explains it. Models are not deep into your specific project, but trained on generic data. They don't really learn on the job, except for what they can feed in their context window.

@allocatooor It's not about that

@dankrad It does feel like it was nerfed from last time

@MaxResnick Claude create a billion dollar business

As long as we can build a benchmark for something, we will eventually be able to build an AI that is better at that thing than Humans are.

How would they milk it? API/usage fees? I think it is ultimately a question who benefits more. Etherscan always did Ethereum for free, and clearly would do more damage to its brand by cutting it off than it could gain from the thread of disabling it. Robinhood could be betting that their chain is so important that Etherscan will do it in order not to lose out.

@0xdoug @severeengineer Once you star paying, you will always pay. Whereas, if Etherscan starts serving it out of their own interest, they will always get it for free.

@severeengineer There is, but it’s blockscout. Whcih is fine, Etherscan is expensive, but like cmon. You’re a $100bn company, you can afford Etherscan

Robinhood Chain didn’t pay for Etherscan. Are you kidding me?

@jonah_b I think there is a lot of demand for holding EUR or GBP, just the people with that demand have easy and familiar ways to do so without using stablecoins.

There is not much global demand to hold non-USD currencies. When given the choice, most people around the world prefer USD. On top of that, non-USD currencies are rarely the primary trading pair in crypto markets, so liquidity is much thinner.

It has been hard for non-USD stablecoins to reach meaningful scale. Here's what I think the issue is...

Today we’re launching tBRL and tMXN. Tokenized Brazilian Real and Mexican Peso, now on-chain. Two of the most liquid + high-yielding currencies globally, tokenized via Tenbin with yield of up to 13% and 7%.

It will be economically advantageous to propose as late as possible. So the natural incentive for a rational proposer who does not try to get a last look (is not malicious) may be to propose so late to sometimes produce partial dissemination certificate. Say social consensus is that it is acceptable to do so 0.01% of the time, then proposers will naturally time their blocks so that they produce certificate 0.01% of the time. A malicious proposer who is trying to get a last look can instead proposer most of their blocks slightly earlier, but 0.01% of the time get a last look. From the point of view of partial dissemination certificates, these two will look the same.

As we argue in the paper, a single such certificate might not be enough, but a repeated pattern of partial proposing can warrant a response by the protocol. There is little excuse for a repeated pattern: as an honest proposer, you freely choose your proposal time (no dependency on a previous block, no view synchronization), so if you observe that your proposals are not reaching enough validators before the deadline, you should simply propose earlier.

Isn't there a free option if I control a proposer and f validators? I can partially disseminate and then delay the fallback consensus if I'm the leader in the first (or even several initial rounds). By that time some of the proposals of the current and even the next rounds were probably decoded, so I can choose whether to publish or not based on this. (With only f/2 attackers this produces a partial dissemination certificate, but I think these aren't very reliable because the incentive for proposers is to publish as late as possible and so they might occasionally produce partial dissemination certificates by accident)

The question is what you do with this information? You can't slash as it is too brittle so it relies on social consensus acting on it, which is hard. What if all reveals are delayed until a validator knows the outcome of all consensus instances on the current and all previous blocks?

Good question. We didn't include this in the paper, but for the option scenario we can show a stronger guarantee than the f/2 certificate, which answers a different question (blaming whoever forced the fallback path). Under synchrony, whenever a proposer keeps an option past the deadline, correct validators can always build a partial-dissemination certificate from f+1 signed No votes plus one Yes vote carrying the proposer's signed commitment, even if the option is never exercised. This holds even with f faulty validators, not just f/2.

@zmanian Because it's the quantity that actually works for most people. (10mg+, on the other hand, seems crazy)

It’s a mystery why melatonin is sold in these doses

@michaeljmcnair @Michael73299790 This is true for all assets however -- if P/E ratios expand and businesses become more expensive, that gain also isn't an increase in national wealth (but the gains due to increase in profits are) The problem tends to be more pronounced for real estate though

@alz_zyd_ Doesn't PPP kind of adjust for this?

Random: I feel like there's an issue with GDP measurement and non-tradables. The plumber in the US does literally the same job as the plumber in a third world country, but gets paid like 100x more for the same job because he plumbs for OpenAI engineers and the other guy doesn't

@iamnotnicola Fair enough! I can now see you can get good context for these local models, that does change things!

@dankrad I think it’s catching up (see glm 5.2), and it is passing the usefulness threshold already (smaller models like qwen 35b a3b are already good enough for daily tasks) It’s unclear you need frontier for all tasks, but I agree that it won’t replace frontier use

in 2026 one needs to decide if they should buy a new car or a personal data center

@flyq_crypto You can still do that, but it becomes 100x easier because every time you don't understand something, you ask your LLM to provide context

@dankrad 我以前一大兴趣/工作方式就是阅读代码。

Fun fact: You don't have to read all the code. You ask LLMs questions about it to verify it does what you want. You make it write tests, or proofs, whatever it takes to convince you that it's correct. This will be more effective than reading the code.