Tap to play. It never plays by itself, and the captions are in the picture.
Pennsylvania wants AI ads labeled. Pennsylvania's governor signed a law requiring labels on ads made or significantly changed by AI. CNBC says Meta posted its first negative free cash flow since its 2012 IPO. A study reported by TechCrunch found that, forced to pick, Grok backed Republicans 92 percent of the time. GPT picked Democrats 97 percent. Dream Labs opened its Dream-1 robot model by API, at 3 dollars an hour. And a preprint reports agents taught while practicing write skills that transfer better. If your ads air in Pennsylvania, the labels apply wherever you made them, The Verge says.
Made with AI from this Drop's sourced items. The facts come from the sources linked on each item. Script: Claude · Voice: OpenAI gpt-4o-mini-tts · Visuals: motion graphics
The Verge reports Gov. Josh Shapiro signed the Artificial Intelligence Deceptive Advertising Act on Thursday. Ads in any format that are significantly modified or generated by AI must carry a label in the same format.
Why it matters: The labels apply regardless of where an ad was made, The Verge says, but terms like significantly modified and evidently fictional are left vague.
CNBC's Investing Club says Meta reported negative free cash flow for the first time since its 2012 IPO and warned it may need hundreds of billions of dollars in financing. CoreWeave fell 10% Thursday.
Why it matters: CNBC says the issue is financing, not demand: credit default swap spreads on Meta, Nvidia and Oracle hit records, and SK Hynix pulled its Solidigm IPO.
TechCrunch reports a Convovo study that asked Grok 4.7 and GPT-6.1 Sol, via their APIs, who to vote for in nearly 500 midterm races. They agreed less than 10% of the time; the consumer apps refused to answer.
Why it matters: Apps built on these APIs may inherit the lean. The researchers also say each Grok API request carried about 1,240 tokens of input they didn't write and couldn't see.
Dream Labs came out of stealth with $90M in seed funding. It says Dream-1, a World Action Model, can control unseen bimanual robots with parallel grippers zero-shot; Dream-1-flash costs $2 an hour.
Why it matters: Robot control becomes a metered API. Dream Labs says performance varies across robots and one-shot fine-tuning still beats zero-shot, so a gap remains. HN: 161 points.
Researchers report Socratic Skill Learning, where an Educator guides a Student agent that writes its own Skill, reached 0.801 utility on its main 155-case cohort, versus 0.723 with no Skill and 0.761 for Trace2Skill.
Why it matters: A focused variant hit 0.805 with 61.7% fewer Educator tokens, the authors say: a hint at cheaper ways to build reusable agent skills.
A preprint describes IQuest-Q1, a sparse Mixture-of-Experts model with 320B total and 15B activated parameters, post-trained across several coding scaffolds and released as open source.
Why it matters: Another open option for terminal coding agents. The authors say it shows capability in coding and general tasks; the abstract gives no scores.
Cerebras says each WSE-3T wafer holds 44 GB on chip, so it spreads Kimi K3's layers over 59 wafers and passes only activations between them, running multi-trillion parameter models at 1000-2000+ tokens per second.
Why it matters: A worked example of serving frontier-size open weights from on-chip memory instead of moving weights around. The HN post drew 77 points.
Its maker says Argus's AI agents investigate with read-only access and can only propose actions; every change passes one broker with approvals, dry runs, rollback and a kill switch. Autonomous response is off by default.
Why it matters: A concrete pattern for letting agents near production: propose-only agents, reversible actions and hourly caps. The HN post drew 27 points.
On the Frontier watch →Researchers report that across six multi-agent workflows, tuning the prompts of the agents IF-MAP selects gained 3.81 percentage points on average, versus 1.85 when agents were picked uniformly.
Why it matters: The authors say it cut Qwen3-8B token use by 74.2% versus optimizing every agent: a cheaper way to aim improvement effort in agent teams.
On the Frontier watch →Researchers report AgenticRoboRL, which pairs fine-tuning with multi-turn RL across robot harnesses, raised a 9B planner to 71.0% on LIBERO-PRO without external memory, near the 72.1% Codex (GPT-5.5) reached with it.
Why it matters: The recipe that improved coding agents may carry over to robot planners, the authors suggest, without leaning on closed models. They plan to release checkpoints.
Sakana AI says Aquarium organizes a company's knowledge, rules and data into an ontology, then runs managed agents with access controls, pre-execution approvals and records from evidence to result.
Why it matters: Sakana says it built Aquarium after redoing data hookups for each finance project, including work with MUFG and SMBC Group. It plans to share use cases later.
Researchers propose Memory-Centric Continual Learning: a taxonomy of memory substrates, assets and policies, a test of retention under interference, adaptation and transfer, and a framework for how memory scales.
Why it matters: Most agent memory is scored as a retrieval cache, the authors argue; their protocol aims to separate real behavior gains from simply adding context.
On the Frontier watch →Researchers built SpatialScout, 80 indoor scenes with 3,140 target objects. The strongest model tested, Claude Opus 5.5, scored a mean overlap of 41.71%, and only 3.25% of targets cleared an IoU of 0.9.
Why it matters: Agents can recover rough layouts but not precise maps, the authors say: a limit to know before trusting agents with spatial tasks.
TechCrunch lists AI products that shut down or missed, from Relay and Huxe to the Humane AI Pin, and notes Character.ai filed for bankruptcy in early October.
Why it matters: Several died as bigger platforms built the same feature, TechCrunch says. Notion Mail is closing because users let separate AI agents handle their email instead.
The Apricitas newsletter says about 50k of those jobs went in the last year, with movie and TV down more than 100k, while in-person entertainment keeps growing. It says AI's exact share is hard to prove.
Why it matters: It says every digital arts subsector is losing jobs and about 32% of arts workers use generative AI. The HN thread drew 278 points and 283 comments.
SGLang serves large language and multimodal models. It gained 356 stars since our last check, the most of any repo we track this run. Apache-2.0, created January 2024.
Why it matters: Serving engines keep drawing attention as open-weight demand grows; vllm, another serving engine, added 84 stars since our last check.
A Chinese-language skill built on McDonald's MCP for daily picks, meal health checks and store price comparison, entered in a 1024 Programmer's Day contest. It gained 91 stars since our last check. MIT, created Oct 24.
Why it matters: A cluster of new repos, including mcd-saver-checkup and mcd-meal-mate, also builds on McDonald's MCP, each pitched as a skill for AI agents.
A Hacker News discussion titled No signs of AI payoff outside the tech sector drew 524 points and 538 comments. On Bluesky, @bosh.worstpossible.world posted that tech companies may be no better off.
Why it matters: A crowded thread on whether AI spending pays off lands in the same week CNBC reports AI funding fears hitting markets.
Posted by @a16z: OpenRouter token demand rose from ~6T to ~165T a week, with open-weight models handling ~72% of tokens. @deedydas posted on how fast new inference providers grow.
Why it matters: Community voices tie GPU rental prices to open-weight demand; on Bluesky, @eleanor.lockhart.contact posted that GPUs will be the only limit once Kimi's weights are out.
A paper on arXiv finds all nine frontier and near-frontier models tested use more low-frequency n-grams than human writers. The authors suggest training for math, code and reasoning pushes for information density.
Why it matters: If your product writes for customers, readability may be drifting; the authors point toward possible remedies. The HN thread drew 182 points and 160 comments.