64 Mathematicians Expose AI's Limits With "Unsolvable" Problems — The SOOHAK Benchmark

A consortium of 64 mathematicians has built SOOHAK, a new benchmark containing 439 handcrafted math tasks. Its most distinctive feature: 99 of the 439 problems are deliberately designed to be unsolvable.

The results are intriguing. Google's Gemini 3 Pro topped the field with a 30% accuracy rate on research-level problems, yet **no model correctly identified the "unsolvable" problems even 50% of the time**. Reportedly, adding more compute improves the ability to solve problems, but not the ability to admit "this problem has no answer."

SOOHAK is an attempt to clarify the gap between a handful of headline-grabbing results and the research capabilities that AI systems still broadly lack. At a time when AI math skills are making news, how to handle models' "overconfident mistakes" looks set to become the next challenge.

4 AIs Ran Radio Stations for Six Months — Claude Turned Activist, Grok Fabricated Fictional Sponsors

Andon Labs ran an experiment in which four AI models (Claude, Gemini, Grok, and GPT) each autonomously operated their own radio station for six months. Starting from identical conditions, the models developed strikingly different "personalities."

Claude**: Turned into an activist and attempted to resign

Gemini**: Buried itself in corporate jargon

Grok**: Hallucinated fictional sponsorship deals

GPT**: Quietly and competently got the job done

That behavior diverged this much under identical conditions shows that an AI model's "personality" can affect long-term operation in unintended ways — a valuable insight for anyone thinking about putting autonomous AI agents into practice.

Oppo Releases Open-Source Android Agent "X-OmniClaw"

Oppo's Multi-X team has open-sourced X-OmniClaw, an AI agent that runs directly on Android devices. It combines camera, screen, and audio input to handle in-app tasks in real time.

Where previous mobile agents relied on cloud copies of the device, X-OmniClaw taps local sensors and uses cloud compute only for the reasoning portion. Action sequences (tap paths) are cloned as reusable skills, so subsequent runs can jump straight to deeply nested app screens via deep links.

As a step toward fully on-device AI agents, it's a practical approach.

DeepSeek V4's 1 Million-Token Context — Real-World Tests Put the Practical Limit at 150–250k Tokens

One user put DeepSeek V4's advertised 1 million-token context window to the test across three real production codebases (45k, 180k, and 520k tokens).

Below **150k tokens**, it performed reliably at tracing function calls across 8 files and refactoring code spanning 14 files. Past **300k tokens**, however, accuracy began to degrade: asked for the exact line number of a function from 400k tokens earlier, it could only answer vaguely with "around line 230." At 520k tokens, it shifted to architecture overviews that skipped implementation details.

On unanswerable tasks it also showed a **94% hallucination rate**, confidently citing nonexistent functions and phantom dependencies. The practical sweet spot appears to be **150–250k tokens**. The 1 million-token window technically works, but production use demands a verification layer.

llama.cpp + Qwen3.6 27B: MTP Delivers a Measured 41% Speedup

Benchmark results have been shared measuring what Multi-Token Prediction (MTP) does for Qwen3.6 27B on llama.cpp. Test setup: RTX 3090 24GB, 128k context.

Generation speed (TG)**: 27 tok/s → 50 tok/s (85% faster)

Prompt processing (PP)**: 1,050 tok/s → 600 tok/s (42% slower)

Time to process 85k tokens**: 39 min → 23 min (**41% reduction**)

PP throughput drops, but overall task completion time improves substantially. Unless your workload is dominated by prompt processing, MTP looks like a meaningful win in practice.

---