AI & design news, merged daily.
AI & design briefing for Thursday, November 5, 2026
Loading today's digest…
No digest available yet today. Check back soon.
Merging & deduplicating today’s sources…
tldr
- Voters approved local data center restrictions, while DuckDuckGo pitched schools a search alternative after Google's AI policy change. OpenAI reported hundreds of Linux exploits and candidate patches and estimated ChatGPT web searches at one-fifth of Google's volume, though its source for Google's figure was undisclosed. Uber reported few AI-agent bookings, and Apple reportedly cut some engineers' AI spending limits.
- OpenAI's policy chief proposed an AI economic agenda, drawing criticism over corporate accountability. Vals found that reasoning costs and performance varied by task, while Prime Intellect published fixes for flawed public evaluations. Linum released an experimental image model trained as a step toward video generation.
- OpenAI introduced GIF creation in Images 2.6, and Fireworks added inference and model-selection options. Claude Code cut the cost of background status updates, while Grok Bot announced phone calling.
Top Stories
- Data center restrictions win in several midterm contests. Local bans and moratoriums passed in several communities, while AI industry and safety PACs saw mixed results. TechCrunch cautions that spending effects are hard to isolate. [1] [2]
- DuckDuckGo pitches schools an alternative to Google's AI search. DuckDuckGo is courting schools after Google said disabling Search's AI features would also disable Search. Google says schools can enable Search separately from Gemini. [3]
- OpenAI reports hundreds of Linux privilege escalation exploits. OpenAI says its research pipeline produced 948 local privilege escalation exploits and candidate patches for human review, at roughly $1.2 million in API costs. [4]
- Uber reports limited AI agent orders. Uber says agents booked four rides and roughly 500 food orders during Halloween week, out of approximately 320 million trips and orders overall. [5]
- Apple reportedly cuts employee AI spending limits. Bloomberg reports Apple lowered some engineers' daily AI allowances from $300 to $200 and is planning further cuts, with exemptions available through senior leadership. [6]
- OpenAI estimates ChatGPT search at one-fifth of Google's volume. An OpenAI product lead offered the estimate, but did not disclose the source of Google's search volume. ChatGPT counts prompts that trigger web searches. [7]
Deeper Cuts
- OpenAI policy chief offers an AI economic agenda. Chris Lehane proposed priorities spanning worker protections, education, safety and infrastructure. An AI Now Institute policy lead criticized the essay's limited corporate accountability proposals. [8]
- Vals benchmarks model routers on completed agent tasks. The new index compares routers and individual models on roughly 1,250 finance, coding, legal and tax tasks, reporting accuracy, cost and duration. [9]
- Reasoning settings show task-specific cost and performance tradeoffs. In Vals' seven-model study, four retained at least 80% of their index score at half the cost or less using their lowest reasoning setting. [10]
- Researchers release fixes for public AI evaluations. Prime Intellect identified unclear tool descriptions and grading issues in public benchmarks, then published fixes. Its revised AutomationBench scores are not directly comparable with Zapier's. [11]
- Linum releases experimental image model trained for $21,000. Linum published weights and code for its Pyramid-JiT image model, describing the compute run as a proof of concept toward larger video models. [12]
- StreamEval examines interactive video models beyond visual quality. An October preprint introduces an open benchmark across five tasks. Its 37-model results suggest visual quality does not guarantee reliable instruction execution or state consistency. [13]
- Researchers propose conditions for shared model representations. An October preprint by Randall Balestriero and Yann LeCun presents a representation theorem for joint-embedding models trained on different views of the same scene. [14]
- A design experiment treats agent skills as applications. The creator of open-source interface-bts describes refining an agent skill's human-facing output, precise instructions and tests across different projects. [15]
- Essay argues for easing existing user behaviors. In an October 28 essay, Rob Ennals argues that product ideas are stronger when they make an activity people already value less painful. [16]
Tools and Launches
- OpenAI introduces Images 2.6 with GIF creation. Images 2.6 arrives in ChatGPT, Codex and the API, with GIF generation and two API model options. [17]
- Fireworks adds inference and model-routing options. Fireworks announced an ultra-fast inference preview, background processing and Nexus Desktop for model selection in existing tools. [18]
- Claude Code reduces background status update costs. Anthropic says status updates cost about 99% less; availability requires version 2.1.316 or later and varies by plan. [19]
- Grok Bot announces phone calling. The Grok Bot account says the assistant can now make phone calls; its announcement does not specify access conditions. [20]
Prompt of the day
One practical prompt for your preferred AI assistant from today's digest. [11]
Build a small, fair test for an AI task I want to use repeatedly. Ask me for the task, available inputs and tools, and what a successful result looks like. Then draft five realistic test cases with explicit instructions, expected outcomes and a simple scoring rubric. Separate genuine task failures from missing information, ambiguous instructions and unavailable tools. Include at least one case where the assistant should ask a question or decline to guess. Make the rubric check the result rather than reward a particular wording, and mark any expected answer that needs my verification. Give me a table I can reuse to compare runs, without claiming the test has already been performed.
References
- TIME: The Winners and Losers of AI's First Big Election
- TechCrunch: AI PACs spent millions trying to shape Congress
- TechCrunch: DuckDuckGo pitches itself to schools amid backlash over Google's AI policy
- OpenAI: A thousand fewer paths to root
- TechCrunch: Almost no one uses AI agents to book their Uber, company says
- 9to5Mac: Report: Apple reduces employees’ AI token allowances as Ternus moves to cut costs
- 9to5Google: ChatGPT has hit roughly 20% of Google Search volume, OpenAI says
- TechCrunch: OpenAI is already trying to shape the 2028 election
- vals.ai: Vals Router Index Leaderboard and Methodology
- vals.ai: How Much Does Reasoning Effort Affect Score and Cost?
- primeintellect.ai: Learnings from Auditing Public Evaluations
- linum.ai: Training a Text-to-Image Model for $21K
- arXiv: StreamEval: Evaluating Interactive Streaming Video Generation Models
- arXiv: Proving and Improving the Platonic Representation Hypothesis with JEPAs
- iamnoman.com: designing a skill
- lesswrong.com: Great but Painful
- OpenAI: Introducing Images 2.6
- Fireworks AI: Fireworks unveils new products at Forge
- X: ClaudeDevs (@ClaudeDevs) on X
- X: Grok Bot (@bot) on X
Sources: The Deep View, The Neuron, AlphaSignal, Superhuman, TLDR AI, Sidebar, The Rundown AI