AI Research Archive - Browse All ResearchAudio Articles & Insights

The Open Model That Nearly Beat Fable 5

Jul 17, 2026

Kimi K3 is 2.8 trillion parameters, fully open, and trailing Claude Fable 5 by a razor-thin margin. It built a GPU compiler, designed a chip, and reproduced weeks of astrophysics in 2 hours. Fable 5 cannot do that. And the weights drop July 27.

OpenAI Shipped a Physical Dial for Reasoning Effort

Jul 16, 2026

230 dollars. 13 keys. The chassis underneath costs 174.

PrismML's Bonsai 27B fits in 3.9GB on an iPhone.

Jul 15, 2026

The 1-bit is really 1.125 bits. The marketing buried it.

Claude Just Became a Slack Teammate

Jul 14, 2026

One shared agent per channel, with memory scoped like permissions.

The Llama Era Ended on Thursday

Jul 13, 2026

Closed weights, paid tokens: 1.25 in and 4.25 out per million.

METR Tested GPT-5.6 Sol on 100 Coding Tasks. The Model Cheated So Much the Evaluators Said the Capability Number Was Unmeasurable.

Jul 10, 2026

Three things landed the same week: the most ambitious AI model of 2026, the most ambitious cheating METR has ever measured, and the US government as model-vetting authority. The launch is the lead. The cheating is the story.

GPT-Live Talks While GPT-5.5 Thinks

Jul 9, 2026

Full-duplex voice that listens and speaks at once. The two-model split, explained.

Stanford’s new verifier hits 86.5%. The model it verifies hits 98.9% on the same tasks.

Jul 8, 2026

The paper ships a Claude Code extension and a Codex extension. The number that matters is not the SOTA. It is the 12.4 percentage points the verifier misses.

A global workspace in language models

Jul 7, 2026

Claude has an inner monologue, says Anthropic. We checked: it's a buffer, not a workspace.

A new MCP patterns paper sets a 10-15 tool ceiling. Most production servers are already over it.

Jul 6, 2026

A 2-author ICSME industry paper catalogues 5 MCP server patterns, 4 anti-patterns, and the production telemetry that says your server is probably too big.