obblekk · HackerTrans

Maybe the most unambiguous "ai will automate work" example I've seen yet.

Absolutely does not imply the workers are automated since they can now use the current models to do more complex tasks at the vast number of new AI training data startups.

Turk was simply not designed for greater complexity tasks and so much of their lunch has been eaten by startups specifically built to collect AI training data.

obblekk

4 か月前議論

Maybe I’m an outlier but I’d rather this than accidentally block legit sites.

Otherwise this becomes just another tool for Google to wall in the subset of the internet they like.

obblekk

5 か月前議論

This is the first time in recent memory that software has had high variable costs so the surprise at these rules is understandable.

In this case, a the difference in context cache hit rate between openclaw and antigravity.

For example if openclaw starts every message with the current time hh:mm:ss at the top of the context window, followed by the full convo history, it would have a cache hit rate if ~0. Simply moving the updated time to each new message incrementally would increase hit rate to over 90%. Idk if openclaw does this but there’s many many optimizations like this. And worse, thrashing the cache has non linear effects on the server as more and more users’ cached contexts get evicted from cache due to high cardinality. The cost to serve difference could be >10x.

Google is the furthest behind on coding agent adoption and has all the incentives to allow off policy use to grow demand. But it would probably be better to design their own optimized openclaw and serve that for free than let any unoptimized requests in.

obblekk

6 か月前議論

One big reason supersonic can be economic now is the increase in wealth in Asia since the 80s.

Transpacific flights from California have no sonic boom population issues for 90% of the flight, and there’s already a large market of people spending $10k on business travel.

Reducing travel time from 12hrs to 4hrs would be a product with a lot more demand than 7hrs to 3hrs to Europe.

obblekk

8 か月前議論

I see 25-29% here https://www.swebench.com/viewer.html for models released in Nov 2024 albeit not verified. gpt4o (Aug 2024) was 33% for swe bench verified.

Important point because people have a bias to underestimate the speed of ai progress.

obblekk

8 か月前議論

80% on swebench verified is incredible. a year ago the best model was at ~30%. i wonder if we'll soon have a convincingly superhuman coding capability (even in a narrow field like kernel optimization).

this is the most interesting time for software tools since compilers and static typechecking was invented.

obblekk

9 か月前議論

1yr timeline is ambitious if it means fully deployed.

Clearly the right thing for Sweden and others to do. Also worrying that even 3yrs into the Russian invasion, bordering countries are urgently increasing their preparedness for future conflicts.

obblekk

10 か月前議論

This is awesome. Coolest hacker demo I've seen in a while.

obblekk

2 年前議論

Human performance is 85% [1]. o3 high gets 87.5%.

This means we have an algorithm to get to human level performance on this task.

If you think this task is an eval of general reasoning ability, we have an algorithm for that now.

There's a lot of work ahead to generalize o3 performance to all domains. I think this explains why many researchers feel AGI is within reach, now that we have an algorithm that works.

Congrats to both Francois Chollet for developing this compelling eval, and to the researchers who saturated it!

[1] https://x.com/SmokeAwayyy/status/1870171624403808366, https://arxiv.org/html/2409.01374v1

obblekk

4 年前議論

Why don't Apple & Google spend a few billion dollars over a few years to rewrite their (non-crypto) unix stack from scratch? It seems like that would be an enduring competitive advantage, good for their users, and reduce future liabilities.

Every programming language can result in bugs, but some are worse/more frequent/harder to solve afterwards than others.

Better yet, why wasn't "rebuild commonly used standard libraries" in the US Infrastructure bill last year? The government could pay programmers a lot, and pay whitehat pen-testers a lot (+ per bug discovered) and in a few years of iteration, we'd have incredibly hardened, durable software infrastructure that would benefit us for decades to come, in the public domain.

obblekk

5 年前議論

this is actually one of the few usecases for which a globally distributed blockchain makes sense.

maybe host an html file on an ethereum smart contract with a very, very large trustfund that people can donate to?

i'd guess the ethereum blockchain will survive 500 years. Or a subsequent fork of it will.

コメント