Local AI on Laptops | AI Field Notes #117

AnalysisAn agent can now be invited into a Webex meeting the way a colleague is. At WebexOne on October 7, Cisco said agents can sit in spaces, meetings and calls and carry out multi-step work across apps, such as pulling usage stats and then updating a PowerPoint deck. Managed agents take on long tasks and come back for approval, and Cisco plans to bring Anthropic's Claude in through that layer. For developers, a Webex Contact Center MCP server (MCP is the protocol that lets AI models call outside tools) is in beta with general availability slated for October, and agent-to-agent protocol support follows in December. Most of the rest lands in early 2027.

AnalysisNothing leaves the laptop, and that is the pitch. Google released Foresight on October 8, an experimental Mac app that transcribes meetings, in person or on any call app, drafts notes and answers questions while running entirely offline on Apple Silicon. It uses EmbeddingGemma 2, a 740-million-parameter on-device model, for capture and a Gemma 4 assistant for chat, and it can search uploaded PDFs, Docs and Office files mid-meeting. TechCrunch frames it as a rival to Granola, the startup note-taker that processes meetings in the cloud. Google shipping a free local tool against paid cloud ones says a lot about where it expects note-taking prices to go.

AnalysisEvery chart arrives with its query attached, and that detail is what makes Claude Dashboards worth a look. The beta, documented in Anthropic's help center this week, turns a plain-language question into a dashboard by writing SQL (the language databases speak) against Snowflake, BigQuery, Databricks, Redshift, ClickHouse or a connected app like Salesforce. It runs on Pro, Max, Team and Enterprise plans, ships switched off for Enterprise until an owner enables it, and counts against usage limits. A companion feature, Claude Motion, turns content into short MP4 animations. Knowing where the tables live was the analyst's moat, and an agent now reads the schema for anyone who asks.

AnalysisUp to 128GB of memory shared between processor and graphics is what lets a laptop hold a large model locally, and it was the spec Nvidia led with on October 7, when Jensen Huang and Satya Nadella launched RTX Spark Windows machines from Acer, Dell, HP, Lenovo, Microsoft and others. Preorders opened that day, laptops ship October 16 and compact desktops follow in November. Microsoft also said MXC (Microsoft Execution Containers) is generally available, letting an IT team set which files and networks an agent can touch on a Windows 11 PC. "We needed to make the desktop the most secure place for agents to execute," Nadella said. The monthly cloud bill now has a rival sitting on the desk.

AnalysisMore than $500 million is what Manus raised after China blocked its sale to Meta, Reuters reported on October 8, with Boyu Capital and IDG Capital leading and Tencent, HSG and ZhenFund joining. Meta had agreed to buy the general-purpose agent startup for about $2 billion last December. Beijing ordered the deal unwound in April under foreign-investment rules, and early backers bought shares back at roughly that price. Reuters puts annualized revenue at around $400 million to $500 million, and The Wall Street Journal reported in September that the round was pitched at a $4 billion valuation. The founders moved staff to Singapore, and Beijing still decided who could buy them.

AnalysisA score of 535.4 out of 600 beats the top human at this year's International Olympiad in Informatics, who scored 498.27, and Nvidia got there with a model anyone can download. Its team said on October 7 that a fine-tuned Nemotron 3 Ultra (550 billion parameters, 55 billion active per answer) sat the IOI live, under contestant time and internet rules, though unofficially and outside the rankings. A separate system scored 30 of 42 at the math olympiad, one point over gold, marked by official IMO graders. Nvidia released checkpoints, training data and a 200-problem benchmark. Olympiad gold has mostly been a closed-lab trophy, and the recipe is now on Hugging Face.

AnalysisSeniors got sharper and juniors only got faster. That is the split in a three-month trial of 133 patent lawyers at 11 firms, published October 7 by Google researchers David Autor and Tanya Rodchenko. With an AI drafting assistant, lawyers' work scored 0.34 standard deviations higher at 10 days and 0.38 at 90. On a final test with AI banned, the AI group still beat controls by 0.32, but the gain came entirely from senior lawyers (0.45). Juniors showed no measurable improvement, though they drafted about 18 minutes faster on a 124-minute task. The tool lifts the floor on output and does nothing visible for the people meant to become the next seniors.

AnalysisOnly 4.5% of eligible US consumers paid for ChatGPT, Gemini or Claude as of August, up from 2.1% a year earlier, according to the seventh edition of a16z's Top 100 Gen AI Consumer Apps, published October 5 using YipitData card data. Spending is lopsided: the top 1% of payers average $903 a month and account for 19.5% of observed spend, while the median payer spends $25. Claude stands out at the top end, with 7.3% of its consumer payers on the $100 Max plan, against 1.3% for Google and 1.1% for ChatGPT. Only 11 new products made the list, the fewest in seven editions. Consumer AI is wide and shallow, and a small group of heavy users pays for it.

AnalysisAltering 0.19% of the weights in one part of a chat model pushed its jailbreak success rate to 53%, while its general test score barely moved, from 52.2% to 51.6%. That is the finding of an arXiv paper submitted October 6 by Muhammad Zeeshan Karamat and Christiana Chamon Garcia and accepted at a NeurIPS 2026 workshop. Working on Meta's LLaMA-2-7B-Chat, they traced safety behavior to a small cluster of parameters in the down_proj layer, part of each block's feed-forward network. GCG, an automated jailbreak method, succeeded 56% of the time. Only one older model was tested, yet an attacker with file access and no benchmark alarm is a plausible threat for models stored on phones.

AnalysisRequests from big power users to plug into the Texas grid went from 63 gigawatts at the end of 2024 to 474 GW by June 2026, more than five times the state's record peak demand, and about 90% of it is data centers. New Public Utility Commission rules took effect October 8: a flat $100,000 study fee and $50,000 per megawatt in financial security, with 20% forfeited if a project loses its slot, at least $10 million for a 1 GW campus. ERCOT (the Texas grid operator) has also paused switch-on approvals for loads over 75 MW, including 17 otherwise finished projects, until an audit due in December. Parking a speculative project in the queue now costs real money.

AnalysisA $4.99 subscription and a few face photos are now enough for an AI to write a first-time acne prescription in Utah. Nolla Health's one-year pilot, approved under the state's AI regulatory sandbox and running from October 5, covers adults with mild to moderate acne and a short list of topical creams and gels. Oral drugs and isotretinoin are excluded. Two licensed physicians approve the early prescriptions before they reach a pharmacy, then review moves to after the fact, then to sampling. Pharmacists are told the script came from AI, and adverse events go to the state within 24 hours. Utah is testing which medical decisions still need a person signing first.