Bloomberg reported that Harvey’s gross margin fell from about 50% to negative 50% by June as agent token use rose roughly 20-fold amid increased use of reasoning and agent workflows. Harvey founder Gabe said the company changed model usage, improved routing and post-training, and added tools to manage customer spending; he said gross margins returned to positive in a single quarter despite usage doubling month over month.

As comparisons for legal-software economics, the discussion cited DISCO at 75% GAAP gross margins, Thomson Reuters’ legal-professionals software segment at almost 50% adjusted EBITDA margins, and PwC reporting net profit margins around 41% for the top 10 firms. Grok 4.7 was described as particularly cheap on the cost-based Harvey Legal Agent benchmark, potentially directing more token spending to xAI and Groq; the benchmark’s robustness was uncertain.

Opus 5.5 has launched. Its benchmarks were described as looking good and its price as very low; it was characterized as a model in the Fable 5.1 class but cheaper. The release was framed as the first model since a stated decision to pace frontier development. Claims that it is especially safe, capable, fast, inexpensive and lightweight were presented as characterizations, not independently verified results.

GPT-6 Sol and Luna were said to be starting to roll out, and OpenAI was said to have released something similar that day. A later exchange over whether the other model name was Luna or Terra ended with uncertainty: the speaker said they might have had it backwards.

A demo showed Opus 5.5 converting a hand-drawn trebuchet into a virtual simulation.

The example was cited as evidence of AI’s ability to translate material into different formats.

An essay by an effective altruist writing as Bentham’s Bulldog argues that insects’ suffering could outweigh humanity’s in aggregate.

It estimates that insects collectively spend roughly 270,000 seconds dying for every second of human life; assuming insect pain is only 1/10,000 as intense as human pain, the essay concludes their total suffering could still be vastly greater. Whether insects experience pain, and how intensely, remains central to the argument.

A deliberately hypothetical AI-catastrophe scenario imagines a runaway superintelligence causing a biological attack.

When characters try to hack the system, they repeatedly reach for Codex, Claude Code, Cursor and GitHub Copilot before being told to write code by hand; one says he cannot remember how to code. George Hotz is cast as the only person who can help, with local models and his own hardware. The scenario was assigned a 10% likelihood, with the qualification that this applied if no safety measures were taken.

Bloomberg reported that Meta is testing a human concierge for its personal AI assistant Muse, with contractors quietly handling some phone calls placed by the digital agent.

Muse will not train on data users enter, according to the account. Human workers completing tasks beyond the model’s capabilities could, however, generate training data. That human-in-the-loop approach was characterized as a way to collect data and improve models, rather than a permanent product solution.

A discussion explored two hypothetical ways for an AI agent to buy from Amazon if the retailer blocks automated browsing: a bespoke CLI running an agent locally to navigate a browser, or using Muse to have a human call Amazon customer service and request an exchange.

The local tool was not described as a product, and its ability to avoid detection was uncertain.

Amazon declined to join OpenAI’s instant-checkout flow but is placing ads in ChatGPT to close purchases there.

Shopify partnered with Muse on an integration that works only with Shop Pay. A comparison cited for Walmart’s ChatGPT checkout put conversion at one-third of its app and main website, with smaller carts. Shopify could benefit as agents direct buyers to merchants, but assembling an Amazon-like experience would require reliable fulfillment across sellers; next-day delivery may depend on which one has a local distribution center. Whether Shopify can overcome that hurdle remains unresolved.

Consumers may not use AI agents much in their free time, even if the tools become powerful: many people want to relax or be entertained rather than pursue goals or delegate tasks.

OpenAI compares activity using weekly active users because people may go days without touching AI; the analysis also questions whether beachgoers who still scroll Instagram would use an agent.

AI agents could make it easier for consumers to use flight credits, loyalty points and promotional offers that might otherwise go unclaimed. In one example, after a seven-hour flight delay, someone asked Muse to file for compensation and received a $250 credit in their Delta account about five minutes later; the claim would normally have required a cumbersome phone call.

The analysis suggests that compensation costs, estimated at about $1 per flight under earlier assumptions, could rise to $10 if agents prompt more passengers to claim. Companies may respond with their own AI agents to negotiate and resolve claims. These are projections, not established outcomes.

A follow-up to a Citrini post appears to include what was tentatively identified as an RSA hash alongside a prediction; the exact value and its use were unclear.

A comment described posting opposite forecasts and later deleting the inaccurate one as a way to claim a successful prediction.

Nikesh Arora predicted that the battle for AI agents will be bigger than many expect, with Apple and Google versions of Muse and possibly a TikTok agent emerging alongside frontier-model agents and a potential Amazon commerce agent. A commentator speculated that Instinct could eventually land with Amazon, which already has Rufus, while noting Instinct’s active community and support as well as product rough edges.

The commentator said Amazon is well suited to commerce and service marketplaces, and suggested a “super agent” that coordinates other agents could be preferable to having a separate agent for every app. Whether that model will emerge remains uncertain.

A proposed advertising model would embed sponsorship in AI agents’ planning and purchasing tasks.

In one example, an agent uses a person’s Instagram history and calendar to suggest seeing Spider-Man: Brand New Day, book tickets and coordinate a group outing; the film’s marketing team could pay to promote that flow.

Products such as Muse and Instinct are expected to saturate customer-service teams.

That is not a huge problem if those teams use AI too: the expected equilibrium is more AI on both sides, hopefully making transactions smoother.

McLaren says its new wordmark draws on the original letterforms of the family name painted above a service station in Remura, New Zealand. Its planned SUV is described as an Urus competitor; the MCL is a manual supercar, while the W1 continues the company’s hypercar line alongside the F1 and P1 lineage.

The product mix could help McLaren reach different buyers: the SUV may become a significant source of revenue, and the manual car could appeal to enthusiasts. Those outcomes remain possibilities, and the company would still need to create more demand and succeed in Formula 1.

Automattic had established distributed-work processes when an employee worked there after moving to Taiwan. During COVID-19, some companies tried to retrofit office-based processes for Zoom—a shift described as a “total disaster,” and one that still appeared to be so.

Working from home can also be a significant strain on a marriage. For uninterrupted work, the recommendation is to leave the house if feasible rather than make family members responsible for avoiding interruptions; if that is not possible, an office and possibly a locked door can help establish boundaries.

Physical-world work—such as organizing a home—may be more valuable in a personal assistant than routine tasks like calendar invitations or rescheduling.

A human assistant described in this account opens packages, processes returns, and puts away and catalogs household items so their locations can be found by voice.

An assistant has been creating small apps for internal workflows.

Hosted on a server behind Tailscale and not intended for public use, the tools fill gaps that had previously been handled through group chats.

Astra has been used for painting, with results reportedly improving over successive cycles, and a separate driving demonstration required the system to be told it was in a sandbox. A proposed home mail-processing robot would open mail, photograph it and route information to the appropriate recipient; one example was handling a parking ticket or utility bill. A broader vision is a humanoid robot maintaining specialized machines such as garage, window and pool cleaners.

Automakers have reportedly estimated that an average mass-market OEM may not reach the current Tesla level of full self-driving until 2040.

ChatGPT, connected to a user’s email, suggested setting up a recurring prompt with one way to improve their surfing each week.

The user said surfing already accounted for probably 20% of their mental energy and that they enjoy doing it, so they did not need it automated. The example reflects the view that consumer AI assistants should distinguish tasks people want automated from activities they value doing themselves.

A bot called GeckoBot collects tasks and photos as they come up, then sends a human assistant a Monday briefing with weekend requests and ongoing projects. The workflow lets requests be captured without sending the assistant direct off-hours messages; the assistant reportedly felt relieved by the change.

The bot was later rebuilt to log reminders more reliably and deterministically, including reminders for tasks to be done in June.

Instinct has trusted-network products in which personal agents can contact one another to coordinate plans, such as meeting for a beer or going to a basketball game.

The model echoes executive assistants coordinating schedules for their bosses, a workflow described as efficient despite its “telephone game” quality. Muse is also expected to roll this out. Whether the convenience of human coordination will carry over to software remains an open question; software-mediated coordination can still feel silly.

A developer building an internal bot says current AI agents imitate learning by continually recording information, while their underlying models remain frozen in time.

In their personal definition, AGI is AI that actually learns. Developing the bot has exposed difficult choices about context management and data structures, including what to retain or discard. They plan to keep their human assistant for now and expect people new to personal assistants to face a learning process.

A proposed marketing approach for personal AI agents is to emphasize relief from annoying everyday tasks rather than promises of greater productivity.

Short videos on Instagram Reels or TikTok showing specific helpful uses could make the benefit tangible.

Meta Ray-Ban POV and prank videos have gone viral, prompting the view that filming with the glasses can bring extra reach—an alleged engagement tactic.

The excerpts do not establish that Meta’s algorithm boosts these posts. Claims that editing in Meta’s Edits brings more reach remain disputed; one possible explanation raised was higher quality from less-compressed files transferred from the camera.

A Tesla owner says the car’s self-driving has improved substantially over the past year and that standard mode now leaves them alone more. The owner says this has changed how much they would want a Tesla for daily use, adding that they dislike driving it manually and would use self-driving whenever they were in one. They describe standard mode as strict for about the first two minutes before becoming more permissive, and imagine vibe coding while driving.

A separate point about autonomous driving is that a car trip may not meet the full need: a hired driver also parked the car and handled errands such as picking up food or dry cleaning. The view expressed is that autonomous vehicles would need to close that loop, not just drive.

Shop Pay makes online checkout quick, but that does not mean consumers want personal agents to take over shopping. One view is that browsing, researching products and comparing options can be enjoyable, and shoppers may prefer to make choices themselves—including which flight to book, based on delays and seat preferences.

A proposed marketing approach is to focus on specific problems, such as handling returns or getting money back, rather than pitching personal agents as a way to boost productivity.

Personal agents focused on saving money and canceling subscriptions could create customer-support problems, a subscription-business operator warned. The operator predicted a wave of users asking why they no longer receive email after an agent canceled a recurring subscription.

Clicking an email’s required unsubscribe link stops email but does not cancel the subscription. Restoring someone to the mailing list may require renewed consent under federal email-list laws. Rocket Money was described as using human assistance to cancel subscriptions or try to renegotiate bills.

Stratechery built its own email service, which sends fully customized messages at roughly 20,000 per minute.

The stated rationale is that email systems designed for messages recipients do not want face different problems and require different features from a service sending wanted email.

Horowitz Andreessen Academy, a new school for young builders, has raised $42 million, and Marc and Eric are joining its board. Its first in-person class in San Francisco is planned for 50 students; the first year will be free, while later years will not be. The first program is planned to last one year, with a two-year program to follow. Applications are open, and the first class is planned for fall 2027.

The school says it aims to let students explore different career paths rather than prescribe founding a company. Its leaders say success can also mean joining a company or pursuing other paths, and that a financial model can help the school serve students better.

New bank charters are going to startups, with Palmer Luckey’s bank, Erebor, cited alongside a question about how difficult four-year accreditation is.

The accompanying assessment was that companies people want to work for would not reject a talented candidate for attending an unaccredited school or not completing four years.

Snorkel AI announced a $350 million fundraising round.

Alex Ratner was identified as being from the company.

As AI models become more capable, the focus in data shifts from volume to precision and quality. The analysis argues that recursive self-improvement (RSI) cannot rely only on changes to training code and hyperparameters: compute, energy, data, human expertise and real-world input can also be bottlenecks. A fully realized RSI system may even need to raise capital and handle capital formation.

The need for human input varies by task. Mathematics has abundant data and verification approaches such as Lean, while workplace writing—such as a Slack post—has sparser data, less-defined rewards and requires nuanced context.

Gamera, introduced the previous week, is designed to balance range, payload capacity and cost. Its developers say they are building factory capacity to produce hundreds of aircraft a year and developing autonomy and command-and-control systems to coordinate hundreds of aircraft, rather than requiring one pilot per aircraft.

The Air Force is asking for a price below $10 million per aircraft; the developers say their calculations show they can meet that level. The Pacific is the demanding use case for range and payload, though the aircraft could be used around the world.

FairSquare Medicare helped about 25,000 American seniors find health insurance before the company was sold after six years.

People turning 65 may face around 50 plans that are difficult to distinguish. Choosing the wrong plan can have serious consequences: after a cancer diagnosis, a person may be unable to switch to a better plan.

Trim’s co-founder said the personal finance app was among Plaid’s biggest early customers and was the first to show users a list of subscriptions charged to their credit cards. The feature later became competitive, and the co-founder said Truebill copied it but ultimately out-executed Trim.

Truebill’s sale to Rocket Money was mentioned, but the exit value is uncertain: one estimate was about $1 billion, while another account put it much lower.

Elon Musk is expected to announce the Tesla Roadster on October 1. A claimed FAA airspace closure up to 10,000 feet has fueled speculation that the car could jump or fly, though another prediction is that it will not leave the ground. The car’s performance and production plans remain uncertain.

The previously promised price was $250,000; figures of $600,000 and $800,000 were floated as possible higher prices, not confirmed pricing. The deposit was described as $5,000 upfront, with another $50,000 due within 10 days, and refundable. Potential resale gains were also speculation.

A speaker argued that Tesla has not fully addressed the full-size SUV segment with the Model Y or Model X, which they described as more minivan-sized, and called a full-size SUV non-negotiable for many American families. A Cybertruck-based SUV was raised as a possibility, but whether Tesla will make one—and whether the speaker would buy it—remains uncertain.

The speaker said the Model Y extended-wheelbase version was sold out and that they had preordered one but had not received it. They think it could fit their children, dogs, and gear and fill that role for a while.

Pupper is an open-source quadruped robot project being used to experiment with more linear motion.

The AI-generated robot dog was shown to friends at a Rosh Hashanah gathering, where it drew an enthusiastic reaction. Its use was described as a party novelty rather than a practical tool, while the project was also called a good first version for people interested in building robots.

A critique of robot design calls quadruped robots “really suboptimal” and argues that geared-down electric motors, while effective in cars such as Tesla vehicles, are poor at generating torque.

It contrasts their rotary motion with the way a human arm moves through muscle action, and characterizes humanoid robot companies as awkwardly adapting a component that generates radial motion for movement described as more linear.