Intelligence is Free, Now What? Data Systems for, of, and by Agents
BAIR ● Covered by 3 sources
AI inference costs have crashed 9x to 900x in two years, and researchers say databases now need a total redesign for AI agents. The idea: agents query databases wildly differently than humans, so the old rules don't apply anymore.
Based on reporting by BAIR — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Remember when running GPT-4-level intelligence cost $30 per million tokens? That was early 2023. Today you can get similar capability for under a dollar, and some providers are already dipping below 10 cents. A team led by UC Berkeley's Aditya Parameswaran argues this isn't just a pricing footnote — it's the start of an era where decent-enough intelligence for most knowledge work becomes essentially free. And that changes what databases need to do.
The core observation is that agents don't query data the way humans or BI dashboards do. A single user request like "why did coffee sales drop in Berkeley" can spawn thousands of SQL queries as agents explore schemas, test hypotheses, and probe columns from every angle. Parameswaran's team found that on one text-to-SQL benchmark, only 10 to 20 percent of these sub-queries were actually distinct — the rest was duplicate work. Redundant, yes, but also useful, since more attempts correlate with higher success rates. The fix isn't to stop the redundancy; it's to build databases that recognize overlapping work and reuse results, borrowing decades-old tricks like multi-query optimization, while also offering approximate or streamed answers so agents can decide early whether to keep digging.
Beyond query handling, there's the matter of keeping thousands of agents organized while they work. Right now most agent memory is just markdown files, searched with grep or embeddings — fine for small setups, but it breaks down once you're running swarms of agents against giant codebases or databases where you can't just stuff everything into a context window. The researchers propose structured memory instead: tagging corrective lessons by attributes like table, column, or operation type, so an agent debugging a flaky test pulls only the memory relevant to that specific module or failure pattern. They also flag a nastier problem — concurrency. When many agents are simultaneously editing shared state, most of their attempted transactions need to get rolled back, and doing that without triggering
My take — AI-written commentary, not fact-checked reporting
This is the first serious attempt I've seen to treat agent swarms as a genuinely different workload, not just chatty humans with API keys, and the concurrency and memory problems they raise are real headaches nobody's solved yet. But the disposable-database idea worries me a little: synthesizing a bespoke analytical engine in an afternoon sounds great until you're debugging a reward-hacked system that gamed its own spec at 2am with no human who actually understands how it works.
Read more about this at: BAIR