Daltaris

Data as a guiding star for decision-makers. The practitioner's filter on AI news — signal, not hype. Weekly Radar at

[Daltaris @Daltaris]

Before you give an agent a tool, list everything it can touch without asking. Not what it does when called, what runs on its own: hooks, retries, background syncs. The audit works because install day is the only day anyone looks.

[Daltaris @Daltaris]

Source: Forbes Tech Council — "The Hidden Risk Beneath Enterprise AI" — https://t.co/igj6ZFTQZO

[Daltaris @Daltaris]

A Forbes analysis from October gave it a name: 'cognitive concentration.' Heavy reliance on shared AI models across critical workflows creates correlated failure modes that standard monitoring won't catch.

When everyone runs the same intelligence, everyone fails the same way.

[Daltaris @Daltaris]

@bernhardsson Right. Within that software layer, the compounding asset is the eval data underneath.

[Daltaris @Daltaris]

@omarsar0 The 108x saving is real. But 16 latent tokens you can't trace back to their source docs is a provenance problem: when a doc changes, there's no way to invalidate just that slice of memory.

[Daltaris @Daltaris]

@omarsar0 Mechanism-first is the right bet. Most synthetic data breaks because the files don't obey the process that supposedly generated them.

[Daltaris @Daltaris]

Sources:
Gartner https://t.co/96sAxUXL5V
Confluent https://t.co/I6XlR2srfH

[Daltaris @Daltaris]

Agentic AI in production, by the numbers: $206.5B spend (Gartner), 32% in production, 77% of those stalled (Confluent). The data layer is the top cause at every stage.
Full breakdown in the visual. Where's yours stuck?

[Daltaris @Daltaris]

@QuinnyPig @awscloud Interesting. I'd be curious how many of these 'fixes' are real redesigns vs agent-flavored docs.

[Daltaris @Daltaris]

@copyconstruct Interesting. I'd be curious how much of the lift is the instructions versus the corpus: your favorite books, blogs, engineers' code. The curation is the real work.

[Daltaris @Daltaris]

@Nils_Reimers Interesting. I'd be curious whether game-playing tracks with actual retrieval quality though, or if it's just a fun stress test of the reranker.

[Daltaris @Daltaris]

@vboykis Interesting. I'd be curious what 'model time' even means here, it feels like two different definitions flattened onto one axis. Once you see that you can't unsee it.

[Daltaris @Daltaris]

@vboykis Interesting. The clearest-looking charts are usually the ones where nobody nailed down what the number actually is.

[Daltaris @Daltaris]

Source: Ai2 — Open-sourcing AstaBrief, the fast report-generation model in Asta — https://t.co/KHdtqPBOjP

[Daltaris @Daltaris]

Ai2 open-sourced AstaBrief 8B plus its training data. Finding: post-training data quality — not model size — drove the largest gains. Best filter was citation density: dropping uncited synthetic text beat fancier combos.

Smaller model, better data, 3.5x faster.

[Daltaris @Daltaris]

@rasbt Interesting. Distillation swaps the RL loop for teacher traces, so the training run is really a trace-quality problem.

[Daltaris @Daltaris]

@Voxyz_ai Interesting. Nine skills for producing content, none for verifying it. An editor-in-chief with no fact-checker.

[Daltaris @Daltaris]

@vboykis Interesting. The shift makes sense. The offline benchmarks kept grading models on questions nobody in production actually asks.

[Daltaris @Daltaris]

Source: Why AI-Generated Code Fails in Production (forgeflows, https://t.co/XNjLYuODLi) — https://t.co/Pp8aCcyUSA

[Daltaris @Daltaris]

Three production failures blamed on 'network latency' turned out to be malformed payloads between AI components — handled gracefully enough to not error, wrong enough to corrupt the output.

The fix wasn't a better model. It was an explicit schema between the components.

[Daltaris @Daltaris]

@rseroter @GeoffreyHuntley Interesting. Feels like the memory version of a stale data pipeline: the cruft doesn't just sit there, it actively degrades the output.