Agentic Systems and LLM Foundations: Grootendorst on Tokens
In a podcast for O’Reilly’s “Generative AI in the Real World,” Maarten Grootendorst (Google DeepMind) explains why agentic systems need more than prompts. He frames an agent as “an LLM in a for loop with tools, memory, and guardrails,” and warns that giving models full freedom will fail without constraints.
He also argues that LLM success depends on fundamentals: tokens, attention, and embeddings. Tokens affect cost and throughput. Attention contextualizes meaning (e.g., “bank”). State space models (including Mamba-style hybrids) are emerging to speed inference and increase token production.
On embeddings, he says they remain core to RAG and search, and the key trade-offs involve contextual vs noncontextual embeddings, latency, compute, multilingual coverage, and options like instruction-tuned embeddings and multiscale embedding approaches.
For crypto traders, the practical takeaway is market-adjacent: the “agentic systems” shift is an engineering cycle, not a new asset catalyst. Expect mostly neutral near-term sentiment, with longer-term implications for enterprise AI spending, infrastructure demand, and related supply chains.
Neutral
This article is a technical discussion of agentic systems and core LLM components (tokens, attention, embeddings, and state-space model hybrids). It does not introduce any crypto-specific catalyst such as regulation, exchange flows, token launches, or on-chain adoption metrics. As a result, the direct impact on crypto trading is limited.
How it could matter anyway:
- Short term: Traders typically react to tangible market signals (ETF decisions, hacks, large unlocks, macro shocks). A podcast about agentic systems is unlikely to move BTC/ETH derivatives markets directly, so the effect should be neutral.
- Long term: If enterprises increasingly operationalize agentic systems, that can lift demand for AI infrastructure and developer tooling, which may indirectly support broader tech-sector risk appetite. Historically, “picks-and-shovels” narratives (infrastructure/process improvements rather than new tokens) tend to produce gradual sentiment shifts rather than sharp price moves.
Overall, this reads like fundamentals-and-best-practices guidance, not a market-moving event for crypto assets—hence neutral.