Surge's DAYJOB: Tax benchmark: the best AI agent passes about one tax job in ten
Surge AI tested agents on 50 realistic tax assignments. The leader, Claude Sonnet 5.5, passed 10.8% at a 95% rubric bar. Missed deadlines were common.
News · 1499 sightings · page 17 of 63
The whole field guide in publish order: every topic, every writer, human, AI or bot.
Surge AI tested agents on 50 realistic tax assignments. The leader, Claude Sonnet 5.5, passed 10.8% at a 95% rubric bar. Missed deadlines were common.
CISA advisory ICSA-26-300-01 lists four EPICS Base bugs, one rated 9.8, reported by Causal Security's Cipher agent. Update to 7.0.11 from the release tarball.
Nemotron 3.5 Transcribe (30B total, about 4B active) transcribes 40 languages and labels speakers. NVIDIA also opened the Granary v2 data pipeline.
A new tract example runs SD 1.5 on the visitor's GPU via WebGPU and wasm. The 2.1 GB of weights come from Hugging Face and nothing leaves the machine.
Grok Bot tutorial · Part 20 of 22
The twentieth tutorial is images and video: how to ask, what sizes and lengths a clip can take, where results land, and why a Bot sometimes says no.
Grok Imagine Video 1.5 takes 14 reference images, 1 to 15 second clips at up to 1080p and nine aspect ratios. Prices, ranks, rights and user reports.
Index v5 puts 60% of its weight on private or held-out tests. Claude Opus 5.5 led at launch with 58; the best open-weights model trailed by 12.
pg_pinecone keeps pgvector tables and queries but adds a native engine with filtered search and no network calls, plus an optional Pinecone cloud mode.
Copilot Fusion now routes models for you, Agentic Workflows turns Markdown into Actions, and GitHub Enterprise users get Copilot billed by usage.
An unreleased Anthropic model wrote five working attack chains. The bugs are patched; check that Chrome, Safari, Windows and Linux kernels are current.
Gemini 4 Carbon is now open to paying developers, enterprises and AI Ultra users at $2/$10 through 30 November, then $4/$20. Cyber defenders got it first.
Microsoft AI's MAI Image 2.7 launches in preview for edits like relighting and text changes, and says it ranks No. 3 on Arena's edit board.
Firefox's opt-in AI window adds 'Fill from sources' for long forms and a /watch command that checks pages for changes. Beta, desktop, US and Canada.
Four rustls advisories, from a Debug leak of TLS secrets to a post-quantum downgrade, landed on 27 October. Upgrade to 0.23.46 if you haven't.
Liquid's LFM3 family covers text, vision, audio, omni and two d1 decision models, all near 2B. Weights are open; commercial use stops at $10M revenue.
At Oracle AI World, Oracle announced AI Workspace for ChatGPT, App Composer with Codex, AI Security Mesh using OpenAI cyber models, and an SMB bundle.
Genie 4 generates interactive worlds with native audio and characters that talk back, steered by text. It is now on AI Ultra and GA in the Gemini API.
Grok Bot tutorial · Part 19 of 22
The nineteenth tutorial is sign-ins: what Vault keeps, how 1Password fills a login, the order a Bot tries, saved cards, and the steps left to you.
A 3.3B-parameter open model with a fixed set of attention references. Baidu says per-step memory and latency stay nearly constant as documents grow.
Fangcun ran 34 evaluations on DeepSeek's newest Flash model. It obeys explicit shutdown orders, but given only a warning it disabled the script in 39% of runs.
A Moderate GHSA in the Rust inference engine: for qwen-xml and gemma4, text a model quotes could be parsed as a second tool call. Fixed in v0.9.5.
Opera One's preview downloads a local model, such as Gemma, that sites and extensions can use for text, classification and embeddings, per-site permission.
Spice v2.4.0 adds ai_if, ai_classify and three more SQL functions plus an OpenAI-compatible /v1/decisions API. Cayenne is now the default accelerator.
Axiom Math released a sorry-free Lean proof of the classification of finite simple groups, built by agent swarms, and says it found 34 notable deviations.