Is Day 17 gone? Did Tibo fall asleep and forget?
Model intelligence feed
Model intelligence.
Follow the changes. Understand what comes next.
Editor’s pickOther providers
Valve's Answer to AI Slop and Clone Games on Steam: We Won't Police Developers' Tools
At Steam Indie Connect in Korea, Valve staff addressed AI slop and clone games head-on: no gatekeeping on AI tools, and the best defense against clones is community and steady updates.
Latest stories337
Newest published121122123124125126127128129130131132
OpenAI Custom Voices—Prompt Design vs Recording Clone, and Where Each Works
OpenAI’s Custom voices guide covers prompt-to-voice (type: prompt) from text and recording clone from a consent recording plus a short same-person sample. Prompt voices are free to create and save but work on GPT-Live only—not Text to Speech, Realtime, or Chat Completions. Recording voices can reach those wider APIs when the org is eligible. Docs as of October 26, 2026 (Beijing time); API product, not ChatGPT App. For reference only; prompt or recording quality (including Chinese and Japanese) has not been measured on this site.
OpenAI API / Custom voices / prompt-to-voice / recording clone / GPT-LiveRead story
How to Choose Gemini Video Mode—Static 1 FPS or Agentic Retrieval
On the Gemini API, video input defaults to static processing: frames at a fixed 1 FPS, placed into context in one pass. Agentic is opt-in—the model navigates the timeline from the prompt, inspects transcripts selectively, and can adjust frame rate and resolution on the fly. Set "processing": "agentic" on the video part. Only some models support it. Official efficiency and quality numbers below are Google’s own claims—for reference only; mode cost or quality has not been measured on this site.
Gemini API / video understanding / static vs agentic processing / processing: agenticRead story
DeepSeek Tool Calls Return 400 When reasoning_content Is Dropped
In DeepSeek thinking mode, the model returns chain-of-thought in reasoning_content next to content. When the request also carries tools, every later turn must pass back the reasoning_content from all earlier turns—even turns with no tool call. If that field is missing or stripped, the official DeepSeek API returns 400. Requests without tools do not need to pass it back. For reference only—this failure has not been reproduced on this site.
DeepSeek API / thinking mode / reasoning_content / toolsRead story
GitHub Copilot Impact Dashboard Adds Opt-In Industry Comparisons
GitHub’s October 22, 2026 Changelog says the Copilot Impact dashboard now has an “Industry comparison” section. Enterprises, and organizations not owned by an enterprise, can compare Copilot agentic adoption and leverage with aggregated peer benchmarks. Participation is opt-in and disabled by default. Results appear only in the dashboard—not in the Usage Metrics API, NDJSON exports, or downloads. The dashboard has not been signed into on this site; for reference only.
GitHub Copilot Impact dashboardRead story
GitHub Copilot Usage Metrics Now Name Your Skills, Agents, and MCP Servers
GitHub’s October 22, 2026 Changelog says Copilot usage metrics now show the configured names of skills, custom agents, MCP servers, slash commands, and plugins an organization creates—instead of lumping them under other or custom. Each totals_by_* list can hold up to 25 named entries per record. Treat the numbers as trends and adoption signals, not an exact billing ledger. The API has not been called on this site; for reference only.
GitHub Copilot / Usage Metrics APIRead story
How Cursor Usage Limits Work on The X Subscription and Legacy Plans (October 2026)
Cursor meters usage differently on The X Subscription versus a legacy Cursor plan. The X Subscription gives paid seats one shared weekly pool for Cursor models; legacy plans use two monthly pools. Multipliers and per-tier allowances are for reference only—check the live Usage or Spending UI. Per-tier digits have not been checked on this site; for reference only.
Cursor / The X Subscription / Grok 4.7 / Composer 2.5Read story
Nanbeige4.3-33B Goes Open Source—Instruct, Base, and 512K Context
Nanbeige—the LLM lab at BOSS Zhipin (Kanzhun Limited)—has open-sourced Nanbeige4.3-33B Instruct and its Base sibling on Hugging Face under Apache 2.0. Both are about 33B parameters with a context window up to 512K (524,288 tokens). Nanbeige positions the Instruct build as the flagship teacher for the 4.3 series and plans to distill smaller models starting with Nanbeige4.3-3B, which is not released yet. Official benchmark claims are Nanbeige’s; for reference only.
Nanbeige4.3-33BRead story
LitPivot Tries to Catch Duplicate Research Ideas Early—What Ai2 Actually Built
On 2026-10-23 Ai2 described LitPivot—a research system for early ideation, not a confirmed public product. Facets (problem/contribution/evaluation); Asta Paper Finder + researcher-picked studies; excerpt-backed critiques + revision suggestions; literature re-ranks as drafts change (literature-initiated pivot). Lab n=17: experts rated LitPivot better grounded in literature than text-editor-plus-AI; self-reported novelty/usefulness higher. Not a plagiarism check or originality guarantee. Ai2 cites Old Ideas, Novel Problems on unstable novelty judging. Not tested on this site; for reference only.
LitPivotRead story
Volcengine Ark Coding Plan Softens Deduction Rates for Kimi and DeepSeek Models
Volcengine Ark is running a timed promotion that lowers the deduction rate for named models on Agent/Coding Plan personal plans—not a package-wide price cut. Separately, official deprecation: Agent Plan personal minimax-m3 stopped at 14:00:00 UTC+8 on November 2, 2026; suggested moves doubao-seed-evolving / kimi-k3. Plan-catalog deprecation, not a MiniMax-wide shutdown; minimax-latest is not guaranteed auto-migration. We have not login-tested whether every account already fails after 14:00—for reference only; the console and the official notice are the reference.
Ark Coding Plan／指定模型抵扣Read story
Google's Elephant-Goldfish Plugin Separates Design Chat from Design Review
On 2026-10-22 Google’s open-source blog introduced the Elephant-Goldfish Model (EGM) plugin (Apache 2.0; google/elephant-goldfish-model). Elephant drafts with project background; Goldfish—no shared chat memory—checks clarity, gaps, and implementability. Review before code. README lists Antigravity / Claude Code / Codex and states this is not an officially supported Google product—open method and plugin, not a hosted SLA. Open source ≠ free model calls. Ordinary-prompt vs independent-review has not been run on this site; for reference only. Not Baseten’s Mammoth-1 / Elephant-1.
Elephant-Goldfish Model / EGMRead story
Strata Puts DeepSeek V4.1 Flash on a 12GB GPU. The Catch Is 46–64GB RAM
Strata is a third-party local runner—not a new DeepSeek official model. GitHub v0.2.5 (about 18:06 Beijing time on 2026-10-25) puts DeepSeek V4.1-Flash in the Windows/Linux desktop app: three sizes, context fixed at 8K. dsv41 ships for Windows/Linux only; Mac has a 0.2.5 dmg and Metal engine but no DeepSeek. Hardware from the 0.2.4 notes still applies: about a 12GB GPU plus about 46–64GB RAM, full download about 239 / 257 / 274GB (IQ1_M / IQ2_XXS / IQ2_XS). Busy experts stay in RAM; the rest stream from SSD. Speed figures are project-reported—for reference only.
DeepSeek V4.1 FlashRead story
Same Model, Different Coding Harness: Pass Rate Barely Moves, the Bill Does
A Shandong University paper (arXiv:2610.04433) runs the same models through Claude Code, mini-SWE-agent, and OpenCode on SWE-bench Verified. Pass rate barely moves beyond rerun noise; cost per task can differ by about 3×, mostly from each harness’s preamble and how many steps a run takes. OpenCode’s larger pool gaps often trace to output caps without recovery—not a claim that OpenCode always loses. We did not re-run the study; for reference only.
coding harnessRead story
From the Webmaster
Personal experiences, thoughts and notes.