AI Search Agents Fail at 'Asking,' Not 'Searching'

A new benchmark called DiscoBench, reported by The Decoder, exposes the real weaknesses of AI search agents in multi-step research. The bottom line: the problem isn't search itself — it's **the failure to ask users follow-up questions when a query is ambiguous**.

The benchmark results show some interesting numbers. Agents that kept searching with the ambiguity unresolved scored 51.9%, while simply answering with a guess scored 43%. Removing the ambiguity from the queries, meanwhile, improved accuracy by up to 40 percentage points. In other words, the biggest bottleneck is that agents "don't ask when they should."

These results suggest that in AI assistant design, the ability to clarify intent through dialogue with the user may matter more than search accuracy itself.

NVIDIA Supplier Hon Hai Posts 40% Revenue Growth — AI Infrastructure Demand Keeps Expanding

According to Bloomberg, Hon Hai Precision Industry (Foxconn), NVIDIA's server assembly partner, reported quarterly revenue growth of 40%, beating market expectations. The company says AI demand continues to grow.

AI rack shipments are expected to maintain their momentum through the current quarter, and information and communication technology products as a whole are entering peak season. Overall business conditions are projected to grow both quarter-over-quarter and year-over-year.

The figures are seen as further evidence that investment in AI infrastructure is not a passing boom but a structural growth trend.

SigMap — A 97% Token Reduction in AI Coding Sessions

SigMap, which has been drawing attention on Hacker News, proposes an approach that cuts context usage in AI coding sessions by 97%. When working with large codebases, context window limits and costs become major practical barriers, and this level of compression makes the approach worth paying attention to as one possible solution.

At this point the project is at the stage of releasing detailed technical documentation, and real-world validation is still to come — but improving token efficiency is a critical theme for the adoption of AI coding tools. Multiple approaches already exist in this space, and how generalizable SigMap proves to be remains to be seen.

Mistral Releases Leanstral 1.5, a Theorem-Proving AI

Leanstral 1.5, released by Mistral on July 2, 2026, is a model for Lean 4, a language used for mathematical proofs and verifying software properties. It has been released under the Apache-2.0 license, with 119B total / 6B active parameters.

On benchmarks, it records scores of 100% on miniF2F, 587/672 on PutnamBench, 87% on FATE-H, and 34% on FATE-X. Japanese-language technical commentary has analyzed the model's value as lying not in "having it write correct code" but in "having the compiler flag its mistakes."

Theorem-proving AI is gradually becoming a realistic option in areas such as software reliability verification and assistance with mathematical research.

Anthropic Begins Building a Severity Assessment Framework for Jailbreaks

Led by Anthropic and others, an effort has started to create a common severity metric for AI jailbreaks (bypasses of safety mechanisms), similar to CVSS.

One of the triggers was a loophole discovered in Claude Fable 5 by Amazon researchers that made it possible to "identify software vulnerabilities one after another." Upon verification, however, less capable models were also able to surface the same vulnerabilities, showing that this is not a risk specific to Fable 5.

This incident drove home an essential question. Jailbreaks tend to be discussed in terms of "did it happen or not," but what really needs to be measured is "how much easier the bypass made things for an attacker." Once a common assessment framework is in place, safety discussions are expected to become more quantitative.

---

Summary

What emerges from today's news is a picture of AI adoption deepening from many angles. Five themes — search agents' conversational ability, sustained growth in infrastructure demand, optimization of coding efficiency, application to theorem proving, and quantification of safety — each show progress at a different layer. What stands out is that all of them connect to a shared concern: not "what AI is used for" but "how we can make AI something we can trust."