Skip to content
Drix10 Blog

AI Systems - LLM Architectures and GPU Compute

, 9 items in Decentralized AI, 5 min read

In this digest (9 items)

LLMs have made tremendous progress in recent years, but their training and deployment still rely heavily on massive GPU clusters. Meanwhile, most developers struggle to access or afford even a handful of GPUs.

Key points

• LLMs require massive GPU clusters for training, but most developers lack access to even a handful of GPUs. • The current GPU compute landscape is characterized by clusters with effectively unlimited compute at one end and startups, researchers, and individual developers struggling to access or afford GPUs at the other. • HarnessDev, a project that puts six frontier models to the test across 2,207 tasks, found that LLM-built harnesses reach human parity in writing and ML, but lag up to 40% in other areas.

Sources

Blockchain - Tokenomics and Staking Requirements

EVAL tokenomics have been updated to account for the current circulating supply of EVAL, which sits at 10 million. The staking requirements have been adjusted accordingly.

Key points

• EVAL tokenomics have been updated to account for the current circulating supply of EVAL. • The staking requirements have been adjusted to account for the current circulating supply of EVAL. • The updates will be outlined in the upcoming EVAL tokenomics on reppo.xyz/token.

Sources

AI Systems - GUI and Intent-Based Interfaces

The GUI has had a 50-year run, but AI can increasingly operate the computer for us. The next interface may simply be intent: tell it what you want done, not where to click.

Key points

• The GUI has had a 50-year run. • AI can increasingly operate the computer for us. • The next interface may simply be intent: tell it what you want done, not where to click.

Sources

Blockchain - Confidential Vault Demo

The Confidential Vault demo shows how ERC-7540 vaults become confidential without sacrificing composability. Encrypted allocations, scoped auditor access, standard workflows preserved.

Key points

• The Confidential Vault demo shows how ERC-7540 vaults become confidential without sacrificing composability. • Encrypted allocations, scoped auditor access, standard workflows preserved. • Try it live at cvault.demo.noxprotocol.io.

Sources

Blockchain - Nscale and GPU Deployment

Nscale has committed $3.5 billion to deploy up to 100,000 GPUs with Figure_robot. Scaling to $6 billion. Brett Adcock says the constraint is compute and data.

Key points

• Nscale has committed $3.5 billion to deploy up to 100,000 GPUs with Figure_robot. • Scaling to $6 billion. • Brett Adcock says the constraint is compute and data.

Sources

Blockchain - CZ and Bitcoin Overtaking Gold

CZ says Bitcoin could overtake gold in the next bull run. His view is that crypto is not a zero-sum game, even for Bitcoin.

Key points

• CZ says Bitcoin could overtake gold in the next bull run. • His view is that crypto is not a zero-sum game, even for Bitcoin. • Bitcoin is only about 10x away from gold, and CZ says it could overtake it pretty soon.

Sources

Blockchain - AMC and Robinhood Tokenized Product

AMC's CEO is challenging Robinhood over its tokenized AMC product. Meanwhile, RH Chain degens have paired that exact token against $MEME, literally "A Meme Coin".

Key points

• AMC's CEO is challenging Robinhood over its tokenized AMC product. • RH Chain degens have paired that exact token against $MEME, literally "A Meme Coin". • As of 4:12am ET, the pool held 547k AMC stock tokens, representing.

Sources

Blockchain - Coinbase and Institutional Adoption

Coinbase is more conservative on shilling memes, etc. I think their bet is to be "the good guys" so they're liked by institutions and the clarity acts moves forward because no way 60 y/o+ decidoors are.

Key points

• Coinbase is more conservative on shilling memes, etc. • I think their bet is to be "the good guys" so they're liked by institutions. • The clarity acts moves forward because no way 60 y/o+ decidoors are.

Sources

Blockchain - GPU Compute and Oil/Electricity Analogy

Carmen Li ( @carmenli ), CEO of Silicon Data ( @Silicon_Data ) says GPU compute is neither oil nor electricity, and explains why on the latest episode.

Key points

• Carmen Li ( @carmenli ) says GPU compute is neither oil nor electricity. • Explains why on the latest episode. • GPU compute landscape is characterized by clusters with effectively unlimited compute at one end and startups, researchers, and individual developers struggling to access or afford GPUs at the other.

Sources