Shopify CEO @tobi explains how an AI agent like Meta's Muse, can grow more valuable to you every day while its maker earns only basis points. AI agents won’t earn 2% of what you buy, only a tiny fraction of 1%. "People are spending money inefficiently... I think the world is not as optimized as people think, and agents can help people save money" However, basis points pay at network scale. for context, Visa earned $40B of net revenue in fiscal 2025, about 29 basis points of its $14T in payments volume. Visa proves tiny fees can make billions, but only because trillions of dollars flow through it An average US household spends $81,905 a year, and 0.29% of all of it is about $235. Your agent can be worth a lot to you while earning its maker about $235 a year from you, at most. The savings stay with you, and the maker needs tens of millions of users to make the math work.

Dean of Valuation Professor Aswath Damodaran just published a long post valuing Anthropic at $1.475T. He says Anthropic’s voting setup, where 7 co-founders hold 50.1% of votes through a single share, “would put Kim Jong Un to shame.” > Anthropic is growing fast but slowing: first-half revenue was about 14x a year earlier, then third-quarter growth cooled as OpenAI won share and Anthropic ran short of compute. > The headline $80.7B first-half loss is mostly a $72B non-cash bond charge; the real operating loss was just over $8B, and Anthropic says it made an operating profit last quarter. > The biggest risk is over $600B in compute commitments, a fixed bill owed even if growth stalls, while founders hold 50.1% of votes, leaving public shareholders little say.

Citigroup now bars third-party AI agents from logging into customer accounts or moving money inside its digital banking platforms, according to the Financial Times. Bank of America has imposed similar limits, joining JPMorgan Chase, which is restricting outside AI agents after Jamie Dimon warned that thousands of “rogue agents” could otherwise operate inside the bank. The deeper threat is to bank funding, since agents comparing rates could move cash from checking accounts paying about 0.1% into fintech accounts paying 3.3% to 5%. JPMorgan, Bank of America and Wells Fargo hold $1.6T in non-interest deposits, and paying 3% on that money would cost $47B a year, close to half their combined annual earnings.

Good news for anyone trying to prove a model is theirs. Sentient Foundation released the Fingerprinting Leaderboard, a free, open source benchmark that tests fingerprinting methods, meant to prove an AI model was copied from yours. Researchers submit a method for hiding an ownership mark inside a model, and Sentient puts that mark into its own test models, tries to strip it out, then ranks the method on a public website. The code that runs these tests, called mlprints, is released too. A fingerprint trains a model to return unlikely answers to secret prompts, so an owner can test a suspected copy through its API alone. But a fingerprint that survives every attack is still useless if the owner can’t read it back reliably. Sentient Foundation’s Fingerprinting Leaderboard scores that verification step directly, alongside how well the mark resists removal and how much model skill it costs. Sentient’s Perinucleus method embedded 24,576 fingerprints in Llama-3.1-8B without degrading utility, is also competing on the leaderboard, and the fingerprints survived standard fine-tuning.

406 researchers, including Yoshua Bengio, Geoffrey Hinton, Stuart Russell and Ron Rivest, signed an open letter today. They want cyber-capable AI models airgapped by default until safer ways to give them internet access are standardized. For models that need outside access, they propose an international body, built with governments and AI and security experts, that would standardize containment practices and track failures, with breaches reported to it and eventually made public.

Ex Google CEO, Dr. Eric Schmidt: AI may hit a money wall before it hits a power wall. Perf per watt had a good run. At 6%, a trillion in debt costs $60B a year. "The real limit to AI is not energy; it is actually cash. When you add up the cost of these things, if you take round numbers, say $50 bn per gigawatt, then 10 gigawatts is half a trillion dollars. How many companies, countries, and so forth can hand an industry a trillion dollars of capital? Very, very few. The Chinese could certainly do it. I do not know if they are doing it, but I am going to try to find out. In America, there are people who hope that is going to happen. It is interesting that you can finance these things because the brilliance of the American capital market allows us to borrow that kind of money. For example, the Europeans cannot do this, which they are sort of sore about." --- Full video from 'Special Competitive Studies Project' YT channel ( link in comment)

FT: Machine-written papers are flooding academic journals, and the cost of screening them is pushing a $22B publishing industry toward mergers. Filtering the flood requires people. thats why Elsevier has grown its reviewer pool by more than a third since 2022 to 1.9M. Those rising quality-control costs, plus fears that chatbots will make paid research less valuable, have kept Springer Nature below its €22.50 listing price 2 years after its IPO. Consolidation is the industry’s answer, starting with Wiley’s £337M purchase of Emerald in June and Informa’s October decision to separate Taylor & Francis, a unit nearing $1B in annual revenue. The same AI that floods journals also pays for them, as Wiley earned about $40M in 2025 from licensing content for model training.

This is beautiful. Grok Bot gains native X search capability. So it can now pull from X directly and work through posts in bulk, instead of stitching together fragments from sources outside the platform. post search, timeline reading, mention checks and activity monitoring. The agent can draft posts as well. Publishing is out of its reach, though. Direct posting has not launched yet, so a human must still press the button.

OpenAI is redesigning ChatGPT replies around tappable buttons, editable graphs and calculators built for a single question, through a feature called Intelligent UI. OpenAI says the goal is making "learning complex topics easier," and users who find the visuals excessive can dial them back like other personality settings. > To make this work, OpenAI built a library of ready-made interface components plus a compiler that displays them as the answer streams, and trained GPT-6 to choose the layout, or plain text when that fits better. > The new model also begins answering before it finishes thinking: OpenAI reports GPT-6 Instant now starts web-search answers 44% sooner than GPT-5.6 Instant, the model it replaces. > They also plan to have ChatGPT generate more of the interfaces people need for everyday tasks, so users describe a goal instead of learning fixed software.

Anthropic dropped Haiku 5.5 > costs 90% less than Haiku 4.5 on prompts up to 100K tokens. > Haiku 5.5 adds the first Haiku effort setting, trading cost for accuracy, but Anthropic still recommends Sonnet 5.5 and Opus 5.5 for complex agentic coding such as Terminal-Bench 4.0 tasks. > Asana reported over 30% lower task-completion latency and up to 2.5x faster inference per agent turn than its current model. > On OSWorld 2.1, which tests whether an AI agent can operate a real computer to finish long multi-step tasks, Haiku 5.5 jumped from Haiku 4.5's 15.7% to 72.4%. On Chartography, a visual reasoning test of reading and interpreting charts without tools, it rose from 6.4% to 46.4%.

Robotics doesn't have an internet-scale dataset. That's the real bottleneck. Mecka just raised a $60M Series B, led by Sequoia, to build one. > custom multi-sensor capture hardware > a computer vision lab turning raw, noisy footage into structured training signal > hiring engineers to scale both A single second of reality holds more information than a page of text. That's the scale of the job.

Most document AI is built to answer one question: where is this number on the page? But the number people need often isn't on the page. It has to be calculated from the numbers that are. Higher-Order Extraction from Unsiloed does that calculation. It calculates numbers a document never prints by generating deterministic code from page values. e.g. you ask for the final figure, it reads the raw values from the document, writes code that combines them, and runs that code to get the answer. Every result traces back to the page. There's a second check. If the document happens to print the answer somewhere, the system compares it with its own result. If they match, confidence goes up. If they don't, confidence drops to zero.

Sam Altman says the public has mischaracterized him as AI's villain for the same reason he once mischaracterized Mark Zuckerberg: when a new technology brings fast change, people project their fear onto the person who seems to hold the most power. admits that in the early days of Meta he held similar belief towards Mark Zuckerberg, at a time when he did not know Zuckerberg personally. ---- Full video on "Fair Game by Vanity Fair" YouTube channel, (link in comment)

This is beautiful. Grok Bot gains native X search capability. So it can now pull from X directly and work through posts in bulk, instead of stitching together fragments from sources outside the platform. post search, timeline reading, mention checks and activity monitoring. The agent can draft posts as well. Publishing is out of its reach, though. Direct posting has not launched yet, so a human must still press the button.

OpenAI is redesigning ChatGPT replies around tappable buttons, editable graphs and calculators built for a single question, through a feature called Intelligent UI. OpenAI says the goal is making "learning complex topics easier," and users who find the visuals excessive can dial them back like other personality settings. > To make this work, OpenAI built a library of ready-made interface components plus a compiler that displays them as the answer streams, and trained GPT-6 to choose the layout, or plain text when that fits better. > The new model also begins answering before it finishes thinking: OpenAI reports GPT-6 Instant now starts web-search answers 44% sooner than GPT-5.6 Instant, the model it replaces. > They also plan to have ChatGPT generate more of the interfaces people need for everyday tasks, so users describe a goal instead of learning fixed software.

Anthropic dropped Haiku 5.5 > costs 90% less than Haiku 4.5 on prompts up to 100K tokens. > Haiku 5.5 adds the first Haiku effort setting, trading cost for accuracy, but Anthropic still recommends Sonnet 5.5 and Opus 5.5 for complex agentic coding such as Terminal-Bench 4.0 tasks. > Asana reported over 30% lower task-completion latency and up to 2.5x faster inference per agent turn than its current model. > On OSWorld 2.1, which tests whether an AI agent can operate a real computer to finish long multi-step tasks, Haiku 5.5 jumped from Haiku 4.5's 15.7% to 72.4%. On Chartography, a visual reasoning test of reading and interpreting charts without tools, it rose from 6.4% to 46.4%.

Robotics doesn't have an internet-scale dataset. That's the real bottleneck. Mecka just raised a $60M Series B, led by Sequoia, to build one. > custom multi-sensor capture hardware > a computer vision lab turning raw, noisy footage into structured training signal > hiring engineers to scale both A single second of reality holds more information than a page of text. That's the scale of the job.

Most document AI is built to answer one question: where is this number on the page? But the number people need often isn't on the page. It has to be calculated from the numbers that are. Higher-Order Extraction from Unsiloed does that calculation. It calculates numbers a document never prints by generating deterministic code from page values. e.g. you ask for the final figure, it reads the raw values from the document, writes code that combines them, and runs that code to get the answer. Every result traces back to the page. There's a second check. If the document happens to print the answer somewhere, the system compares it with its own result. If they match, confidence goes up. If they don't, confidence drops to zero.