In an interview, Nvidia CEO Jensen Huang argued that AI development need not be paced simply because labs face difficult safety challenges. If a problem can be solved through engineering, he said, its cause should be identified and processes improved; if experiments cannot be contained and could damage the world, he argued, labs should be shut down. The interview framed competitive pressure among AI labs and with China as a concern that could push development too fast.

Huang used self-driving cars as an example of a product that should not be shipped when its safety problem cannot be solved. Separately, a commentator suggested courts might hold a company liable if its AI caused economic harm, while noting that a regulatory gap could exist.

The AI effect on employment appears “pretty minimal” so far, according to figures cited in an analysis: unemployment among U.S. white-collar workers was put at about 3%, while Philippine unemployment was cited at 4.9% in June 2026 and 6% in August. Call centers account for about 3.5% of the Philippine job market, against a previously cited warning that 90% of the country’s economy could disappear because of AI’s effects on the sector. The analysis cautions that impacts may take longer to emerge and could look different at a much larger scale.

Jensen Huang is characterized as viewing AI as a normal technology. The analysis compares AI with the internet, which transformed daily life and created wealth but did not, in this view, produce an obvious jump in broad economic data. AI may instead raise output by speeding up many existing tasks: efficiency gains can lead people to do more work in the same amount of time, and faster transactions could support economic growth even without creating entirely new jobs. The long-term effects remain unsettled.

The 10-year bond yield was described as above 5%, its highest level in 19 years—not an all-time high.

The increase was presented as widely attributed to the war. Commentary questioned whether higher yields were worth celebrating, joking that cheering a bigger number because it starts with 5 could overlook its implications.

Reported examples include Claude Opus 5.5 generating every frame of an animation in JavaScript, and a voice-prompt workflow that turned a children’s story into a series of stories and books, sought a printing option, and requested illustrations, a short-form video and a video game. The latter was described as using subagents. Another example converted an AI-generated image into an SVG of about 40 MB.

The proposed direction is for models to select tools such as Slack, Photoshop or Python as tasks require. The view is that the best tool and the cost of getting the result matter more than synthetic benchmarks of code cost; creative work can also move between formats, such as Blender, After Effects and PNG.

A report said Ahmed al-Sharaa was spotted scrolling and sending Instagram Reels during the UN General Assembly while wearing translator headphones.

It was not clear why he was using the feed. One possibility raised was that a personalized stream of local political content could help him gauge public opinion and what matters to voters ahead of the next election; this was speculation.

A founder of an AI film-production company was described as doing very well and hiring many video editors. The reported pace was unclear: “4 video editors like a week or a day or something.” The work involves selecting and processing generated output, and keeping style and pacing consistent—especially in larger projects, where consistency matters more over two hours than in a one-minute piece.

Models have become cheaper since concerns about negative margins at Harvey were discussed, prompting a renewed question about whether value will remain in specialized applications or whether companies will work directly with foundation-model labs. That shift remains a possibility, not a settled outcome.

One proposed, explicitly arbitrary economic yardstick defines AGI as the point when AI revenues match non-AI revenues—or when the AI economy is as large as the human economy. The measure was described as interesting, though its value was questioned.

The accompanying estimate put AI’s contribution at about a quarter of a percent of GDP. Total AI-lab revenues were described as being in the hundreds of billions, compared with an economy measured in tens of trillions globally or even in the U.S.; by this comparison, the threshold has not been reached.

With the Tesla Roadster reveal said to be about a week away, an unverified rumor relayed by someone who claimed to have sat next to a Tesla employee says the car’s jets are meant to generate downforce, not make it fly.

The same account predicts 0–60 mph in one second; the source cautioned that the person could have been joking.

Sion CER Exobot is described as Saudi Arabia’s first true production-car brand, with a sedan and an SUV planned. A vehicle was also described as having an 850-horsepower, tri-motor electric powertrain; the specification was not tied to both body styles.

No price was specified. A price in the hundreds of thousands of dollars was raised as a possibility, while six-figure EVs were described as difficult to sell.

The trailer shows Nathan Fielder proposing to move into Elizabeth Holmes’s home before she goes to prison and document the process. Its dark, unsettling tone has prompted speculation about how the film will portray the Theranos case.

One prediction is that the film will go beyond the straightforward account of Holmes defrauding investors and failing to build a reliable blood-testing device, perhaps complicating assumptions about her guilt or suggesting another person bore more responsibility. The trailer does not establish that such a twist occurs.

At the Snapdragon Summit, Qualcomm said it was announcing its latest flagship smartphone processor and described smartphones’ shift toward becoming AI devices as already beginning. The company said it had previously demonstrated a multi-billion-parameter model running on-device.

Qualcomm argued that asking whether AI should run on the device or in the cloud is the wrong question: apps will use both, with the split intended to be transparent to users.

Amon says Qualcomm now serves car companies around the world, with on-device AI already playing a major role in assisted and autonomous driving. The AI runs on a processor in the car, he says.

He sees agent experiences as a natural fit for cars, which do not carry the legacy of operating systems and apps found elsewhere. Cameras and other systems designed for navigation could also let an agent answer questions about places in view—for example, a restaurant’s Yelp reviews and whether it has lunch availability. Amon says this convergence is accelerating for Qualcomm.

Compute demand currently exceeds availability across the board, while the semiconductor supply chain is operating at 100% capacity. Data centers are expected to evolve toward heterogeneous computing, with different processors or engines handling different tasks.

CPUs are expected to remain important for orchestrating agents, and CPU demand is projected to continue rising in data centers and on devices as agents are deployed. In inference, separate engines may handle stages such as prefill and decode.

Amon said most new classes of devices are already using Qualcomm chips. He expects smart glasses to reach an inflection point, arguing that their cameras, proximity to the eyes, mouth and ears, and head movement could enable functions such as “see what I see, read what I read, hear what I hear.”

He described glasses, camera-equipped earbuds, jewelry, pins and buttons as emerging personal AI device form factors, and predicted that the category will become very large.

Qualcomm said it acquired Modular, whose stack it described as similar to CUDA but designed to run on CPUs and accelerators from NVIDIA, AMD, Qualcomm and other hardware. Qualcomm identified Chris Lattner as Modular’s founder, with the qualification “I think,” and said he created Apple’s Swift programming language and LLVM.

Qualcomm said it is making the stack open source and believes an open platform that scales from data centers across different hardware to the edge could benefit the industry and drive AI innovation. It said it would support the stack even among competitors, arguing that otherwise one company could do most of the innovation.

A CEO said discussions in the United States often focus on model safety, with cybersecurity a particularly serious issue because AI has a large attack surface. The CEO also said products should be developed responsibly.

In China, the conversation is different, the CEO said: it focuses on putting AI in every car, phone, PC and industrial setting. The contrast reflects a wider gap between discussion of AI’s practical deployment and its broader implications.

A prior outlook held that AI models would get cheaper and that short-term margin compression could eventually iron out, rather than signal a collapse.

The counterpoint was: “Your margin is my opportunity.”

Fall coined “token market fit” for areas where an average end user can productively use about $10,000 a month in tokens. The areas identified as showing this fit so far are coding, video and media creation, and possibly high-frequency trading. Physical AI was also named as an area of interest.

The assessment said legal work, customer support and sales had not yet reached product-market fit “in a real way,” while predicting they would. It also suggested bottlenecks were moving away from coding and video.

Bessemer Venture Partners has long invested in Legora, an example of legal AI.

The product is still a copilot in many use cases, rather than an autopilot doing lawyers’ work, and large law firms have not yet massively changed their team compositions. More automation is expected, with spending on agents in legal fields potentially exceeding spending on models.

If billing moves into AI tools such as Claude Code or Codex, those platforms could take an app-store-like cut.

It remains an open question whether that would hurt inference providers or whether increased distribution could offset the fee.

A small seed investor in Instinct called personal AI agents a potentially major next category in AI, saying the products had made a strong impression even on family members who do not use technology. Unlike the Claude Code and Codex boom, which especially appealed to people comfortable with terminals and code, these agents can offer value to people without a GitHub account.

The investor expects major AI labs and companies to compete to create similarly appealing agent experiences and bring them to enterprise use. She said their simple interactions and proactive features could transfer to enterprise, and predicted a market with multiple agents and approaches rather than a winner-take-all outcome.

Bessemer said it raised $5.75 billion, with $1.75 billion dedicated to early-stage companies and $4 billion for growth investments.

The firm said about 70% of its investments began early, sometimes before a company had revenue or even a name.

Companies are staying private much longer, while trillion-dollar IPOs—unthinkable years ago—are now part of the landscape.

Real is a score and statistics app for fantasy players, bettors and sports fans. It turns live play-by-play data into a readable, contextual feed with comments and reactions, including milestones such as career achievements and scoring streaks. Real describes the underlying data as widely available: Genius Sports and SportRadar scrape much of it, and a basic fact such as Steph Curry hitting a three is not owned by leagues or others. Low-latency data is especially important to gambling companies.

A cited figure put the app at “1, 1.5 million” monthly active users; 30% of monthly users reportedly leave comments. Chats are organized around specific box-score events and player performances rather than one broad game chat.

Real’s first 1,000 users were friends who ran meme fan pages, some with hundreds of thousands to millions of followers. The app began with NBA coverage and expanded to 18 leagues, with creators encouraged or paid to use Real scoreboards in game-related content.

Real also turned the top three stats from a game into graphics designed for sharing. Users share or screenshot stats to praise or criticize players, a practice described as helping the app grow.

Real’s founders described a virtual-currency model in which users can earn currency and collect digital representations of individual plays—such as home runs, three-pointers and touchdowns—through real-time raffles or general packs.

The app rates performances from 0 to 10 and can score above 10; its founders said users refer to the ratings in rookie rankings and MVP discussions, and some players check the app during games.

Real sends alerts for notable live sports moments, such as a player’s scoring streak or a touchdown, and selectively highlights baseball players nearing a cycle—for example, alerting users when the remaining hit is a single, double or home run rather than a less likely triple. The average daily notification figure was stated as “50 50 million.”

Real’s alerts were contrasted with ESPN and Bleacher Report score apps, characterized as generally sending updates at halftime or the end of a game. The app was also described as a fast, text-based way to follow multiple games at once.

Real is starting to add golf and tennis, with F1 planned, alongside existing coverage of soccer, the World Cup and karate. The product is considering a community vote on obscure sports; AFL is repeatedly requested by its substantial Australian fan base, including daily requests around 3 a.m.

Demand is cyclical, and Olympic basketball coverage is under discussion, though it would require engineering time for an event described as taking place for four months of the year. Real’s updates have accelerated from every three to four months to every two to three weeks, which the company attributed to increased engineering output with AI. It has added next-generation NFL statistics, including telemetry and personnel and coverage details, and says most usage is live.

Real said it launched a virtual-bucks feature for users to build profiles and compete, and that fantasy players and bettors accounted for much of its usage over the previous year and a half to two years. The company said it had considered prediction markets and discussed real-money gaming options including Kalshi and Polymarket before choosing FanDuel to bring odds into its product, calling the partnership another revenue stream. It said live odds could add data and discussion to the product, with an option to turn the display off.

A related assessment characterized Real’s engagement product as potentially valuable to betting and prediction-market companies, citing user acquisition and deep usage, and suggested it could be difficult to replicate.

A product is positioning profiles built around fandom and viewing as a way to create a deeper community than simply spending time on Instagram, Twitter or TikTok.

Its team wants to strengthen that community dimension. The potential spans niche interests, with hundreds of millions of Taylor Swift fans cited as one example.

Sports may be relatively resilient to AI because live experiences are among the few human things people still value.

Their appeal was linked to rising ticket prices, and tickets were also described as expensive; no evidence was offered to establish that demand for live events is driving the increases.

Real says it has used AI for moderation since day one and now scans every reply in real time. Its process starts with a lookup table that catches an estimated 80–85% of problematic content, followed by a fast classifier that catches another 10%; anything that gets through is checked by newer models, with escalation when needed.

The company describes moderation as a balance between blocking seriously harmful content and retaining some edginess. It says the system has worked well and argues moderation is important for live chats, some of which otherwise have to be closed.

A fantasy app built around The Bachelor, with its creators manually logging events such as kisses and hugs, had about 50,000 weekly users. Its most-used screen was a live feed where users commented and watched along as a second screen.

The product later peaked at about 1.1 million daily users. During a Bam Adebayo game, around 200,000 people simultaneously tracked each point as he approached 83 points. Its team identified more sports, content and in-product data, as well as live events ranging from The Bachelor to Love Island, as possible ways to pursue a billion users. It also called making major-event use more engaging and fun with friends a potential growth unlock.

People often ask Real about elections, but the company said it does not want to go down that path. It said sports still has “like a 10x” potential with its current offering.

An election map filling in red and blue overnight was described as resembling the visualization of a prediction market.

Real’s app lets users create groups nested under any piece of content, where members can talk. Many subcommunities are still global and form around unexpected meme or sports moments, such as people following Brandon Podziemski’s attempt to score 30 points.

The team says it does not create these moments. As the product grows, it thinks it may begin curating such groups or encouraging them beyond communities centered on teams or players.

A Geely electric vehicle was described as able to charge in under five minutes, a speed viewed as a possible tipping point for wider EV adoption. The possibility of new, faster charging technology for the Roadster was also raised, without specifics.

The Porsche Cayenne EV was cited as offering wireless charging via a mat on a garage floor. Tesla was noted as not having moved quickly on wireless charging. Charging at stations can cost tens of dollars and takes considerably longer than refueling with gasoline, making shorter charging times important, the source said.

Affirm announced that its service is now available on Amazon in the UK, extending a partnership the companies have had in the US for some time.

Affirm is also live in the UK with Costco and previously expanded there with Shopify.

Affirm says it launched a suite of attention-driven underwriting models a few days ago.

The company says it has built underwriting models for 15 years and trained the new models on a large volume of data. After primarily using tree-based models, which it said are easier to audit and explain to regulators, Affirm began researching attention-based approaches about three years ago.

Affirm set a public target to become profitable 24 months in advance and reported progress quarter by quarter, saying it reached profitability a couple of months before the target. The company’s account described public reporting as a credibility benefit, but noted that public companies can face market penalties for missing metrics. Private-company metrics, even when published, are not communicated under the same GAAP and SEC-regulated standards; venture capitalists are described as the people with intimate access to those internal figures.

Execution gaps can take time to become visible: the example raised was two vertical AI companies with $500 million each in funding, whose relative performance might remain unclear for roughly two years. As possible early signals, command of company metrics and the full technology stack was cited; engineering-led companies and leaders with engineering degrees were suggested, not guaranteed, to be more likely to follow through and tell the truth.

Affirm’s CEO predicted the company would have a role in agentic commerce, arguing that financial specialists will handle underlying payment and credit choices as agents take on some tasks. He said people will remain involved in some shopping decisions, including how to pay, and that Affirm is working to improve underwriting so it can serve more people while maintaining credit performance.

He said the company must move faster and differentiate itself from competitors, arguing that a smaller, less profitable version of Affirm would not displace it.

Affirm formed a developer-productivity team of about 12 people around January to curate AI models and coding tools for engineers, including Cursor and Claude, and make it easier to move to better options as they emerge.

The company says the team keeps its tools near the cutting edge of commercial and open-weight models and coding harnesses.

A company that raised $100 million said the funding will be focused primarily on research and development.

Insight Partners and Salesforce Ventures were named in connection with the round.

Starting in January, all businesses selling software in California have to collect and remit tax on software.

No year was specified.

As AI takes a larger share of the economy, more tax revenue is expected to come from taxing AI-related activity.

AI has penetrated tax advisory less than fields such as law, an assessment holds, leaving potential to build a very large business.

The company uses several customer-acquisition channels, including partnerships with firms, direct sales and marketing.

Partnerships are becoming more important as it moves upmarket, where complex businesses trust their advisers.

An unnamed company says it focuses on indirect taxes, including sales tax and VAT, and files in more than 80 countries. It aims to ease the low-level work handled by tax workers so they can focus on more strategic tasks.

SponsorRamp is also building in the space, but the company says it does not view SponsorRamp as a competitor, describing the market as broad and divided into many sub-verticals.