Artem Zakharov @pacary

Joined July 2008

Tweets

2K
Followers

2K
Following

827
Likes

2K

Artem Zakharov @pacary

a month ago

4/4 And most likely, it will solve that problem in the most efficient way possible. A new world.

0 0 0 100 0

View Details

3/4 Because right now we are living in a time when a single person — or a handful of people — can launch a million or even millions of GPUs so that an LLM can think about a single problem for virtually any amount of time.

2 0 0 123 0

View Details

Artem Zakharov @pacary

a month ago

1/4 Superhumans. Overhumans. It was a kind of concept I first encountered — without really understanding it at the time — when I read Professor Dowell’s Head by the science fiction writer Alexander Belyaev when I was 12 or 13 years old.

1 0 0 151 0

View Details

Artem Zakharov @pacary

a month ago

2/2 Now LLMs can do this work almost for free and hundreds of times faster than humans. In other words, the commercial open-source model is practically dead and will be fundamentally rethought in the very near future. And it may eventually evolve into something entirely new.

0 0 0 52 0

View Details

Artem Zakharov @pacary

a month ago

About open source. 1/2 There are many great companies that were built as commercial businesses while keeping their core product open and free. Most of their revenue came from integrations and custom development for clients.

1 0 0 73 0

View Details

Artem Zakharov @pacary

a month ago

2/2 You can create multiple Docker containers, where each Openclaw/Hermes instance acts as a separate agent dedicated to its own task or set of tasks, without access to the main system. Even if something goes wrong, it can only affect its own isolated environment.

0 0 0 90 0

View Details

Artem Zakharov @pacary

a month ago

1/2 Openclaw/Hermes are useful tools. But you also shouldn’t give them full access to your entire computer. To make sure they work properly without breaking anything, ask ChatGPT/Claude to deploy them inside Docker — basically an isolated computer inside your computer.

1 0 0 114 0

View Details

Artem Zakharov @pacary

11 months ago

@jack There’s a serious limitation: TestFlight mode restricts the number of installations, which makes it unsuitable for a functioning network. Without sufficient reach, exponential user growth becomes impossible.

0 0 0 28 0

View Details

Artem Zakharov @pacary

a year ago

2/2 One way to slightly improve the output is to ask the model to add excessive logging to all code. That way, it can better understand what’s happening in the moment while fixing individual errors.

0 0 0 328 0

View Details

Artem Zakharov @pacary

a year ago

1/2 LLMs write tons of code per minute. But when you try to run the resulting zoo, most of it doesn’t work. Then the LLM starts rewriting the code—and so the cycle begins, looping endlessly until your eyes start bleeding. Or until your account runs out of money.

1 0 0 348 0

View Details

Artem Zakharov @pacary

a year ago

2/2 That is, investing $1000 in LLMs yields somewhere around $50,000-70,000 in human gross product just in terms of costs. Obviously, there are many caveats and "what ifs" here.

0 0 1 186 0

View Details

Artem Zakharov @pacary

a year ago

1/2 Labor productivity is quite a complex category when we move away from measuring in units, such as brick production. I tried to subjectively assess the labor productivity of current, most productive LLMs. By my most conservative estimates, I get approximately x50-70.

1 0 0 344 0

View Details

Artem Zakharov @pacary

a year ago

@AndreCronjeTech It looks like an 'operating system' for DeFi 2:0, where the base AMM + RWAP provide all the necessary metrics for other components to operate

0 0 1 39 0

View Details

Andrej Karpathy @karpathy

a year ago

DeepSeek (Chinese AI co) making it look easy today with an open weights release of a frontier-grade LLM trained on a joke of a budget (2048 GPUs for 2 months, $6M). For reference, this level of capability is supposed to require clusters of closer to 16K GPUs, the ones being brought up today are more around 100K GPUs. E.g. Llama 3 405B used 30.8M GPU-hours, while DeepSeek-V3 looks to be a stronger model at only 2.8M GPU-hours (~11X less compute). If the model also passes vibe checks (e.g. LLM arena rankings are ongoing, my few quick tests went well so far) it will be a highly impressive display of research and engineering under resource constraints. Does this mean you don't need large GPU clusters for frontier LLMs? No but you have to ensure that you're not wasteful with what you have, and this looks like a nice demonstration that there's still a lot to get through with both data and algorithms. Very nice & detailed tech report too, reading through.

DeepSeek @deepseek_ai

a year ago

🚀 Introducing DeepSeek-V3! Biggest leap forward yet: ⚡ 60 tokens/second (3x faster than V2!) 💪 Enhanced capabilities 🛠 API compatibility intact 🌍 Fully open-source models & papers 🐋 1/n

649 2K 13K 7.4M 5K

391 2K 19K 6.5M 8K

View Details

Artem Zakharov @pacary

3 years ago

3/3 An incredible leap, all within just 5 years. More to come! ❤️

0 0 3 307 0

View Details

Artem Zakharov @pacary

3 years ago

2/3 It seems 'every dApp' is now crafting its own network or rollup. Those who've built something functional often find real users are scarce, with most being 'degen-retrosearchers'. The level of expertise has skyrocketed, with almost no fluff in personal talks.

1 0 1 319 0

View Details

Artem Zakharov @pacary

3 years ago

1/3 After ETHGlobal Istanbul: There's a bit of confusion among participants. They're diving into micro-use cases and eagerly jumping on the zk bandwagon, a term that's been everywhere. RWA discussions are surprisingly sparse.