More and More Websites are Becoming No-Go Areas for Agents

Yelp, Delta, JPMorgan, Citi, and Bank of America: even more websites are blocking agents

Several major US websites are currently blocking autonomous AI agents designed to perform tasks on behalf of users. Yelp stated that it only allows agents that pay for its in-house data licensing program, while Delta Air Lines prevents agents from booking tickets. JPMorgan Chase and Charles Schwab confirmed that they block agents from accessing customer accounts for security reasons. In recent weeks, users have reported blocks on Meta’s Muse, xAI’s Grok, OpenAI’s Dots, and Instinct’s agent; Cars.com and Amtrak have also suddenly stopped working for some, suggesting that other providers are quietly implementing blocks. Even without explicit bans, agents regularly run into anti-bot systems.

Amazon set the precedent last year by blocking Perplexity’s shopping agent and subsequently filing a lawsuit. Amazon argued that it had repeatedly asked Perplexity to stop, and that third parties shopping on behalf of customers should operate openly and respect the service provider’s decision. Perplexity countered that Amazon simply feared fewer human eyes on its ads and compared its own agent to an extension of the user. A court initially prohibited Perplexity from shopping on Amazon, an appeals court overturned the decision in August, and Amazon then filed an amended lawsuit.

Meta’s Muse recently received the same message from Amazon: continued access violates the site’s policies. Elon Musk commented on the blockades by suggesting he might have to start an Amazon competitor. In a post, Meta stated that whoever rejects a personal agent is rejecting the customer behind it, and is opting for partnerships with Walmart, Gap, Sephora, and Best Buy instead of confrontation; Shopify has also opened its network of shops to agents. Chris Canal, CEO of the verification company EquiStamp, said the blocks damage the perceived usefulness of the providers because agents can do less than promised. OpenAI and Google did not comment on request, and Instinct declined to comment. → The Block, The Verge, Wall Street Journal

Synthszr Take: Dimon’s phrasing of “liability transfer” is the real core of the block: as soon as an agent logs in with your credentials, the bank sees a regular customer and can fall back on the terms of service in the event of a loss. With real-time payments via Zelle, the money is gone, and to this day there is no rule that specifies who replaces a transfer triggered by a third-party agent. At Wachovia, 20 percent of call center complaints came from scraping traffic, and those were systems that only read; agents act.

Anthropic is having lawyers write negotiation rules for Claude agents

Anthropic is working with law firms on standards for how AI agents can negotiate on behalf of users. According to the Financial Times, this involves agents that negotiate prices, handle complaints, or renegotiate contract offers. In its own experiments, the company has shown that more powerful models systematically achieve better deals, which could increase social inequality. To this end, the company is already recruiting so-called Conceptual Reasoning Fellows, specifically lawyers, who will test Claude for four months on judgment-heavy questions from law, philosophy, cognitive science, economics, and political science. The compensation is between $100,000 and $167,000.

The background is that autonomous agents are encountering more and more closed doors on the open web. Users of Meta’s Muse, xAI’s Grok, OpenAI’s Dots, and Instinct are reporting blockades on major websites. Yelp only allows agents that pay for its data licensing program, Delta Air Lines prevents ticket bookings by agents, and JPMorgan Chase and Charles Schwab are blocking account access, citing security concerns. Cars.com and Amtrak are also no longer working for some users. Instead, Meta is focusing on partnerships with Walmart, Gap, Sephora, and Best Buy, and Shopify has opened its merchant network to agents. → The Verge, Crypto Briefing, The Decoder

Synthszr Take: With a reported valuation of $965 billion, negotiation standards from law firms are the cheapest item in the preparation for an IPO. Governance here is a capital market product: Anthropic is selling auditability to investors who, before subscribing, want to know what liability is contained in an agent that negotiates prices and handles complaints on behalf of millions of users.

Microsoft builds Surface Laptop Ultra with Nvidia chips

In an interview, Microsoft CEO Satya Nadella unveiled the new Surface Laptop Ultra, which is based on Nvidia’s RTX technology from the Spark platform. In the same conversation, according to TechRadar, Nadella outlined Microsoft’s plan to make Windows an operating system on which AI agents can run more securely. Nvidia CEO Jensen Huang is also quoted in the report, describing the jointly developed device as “really a work of love”; he said the computer will be “incredible.” Both assessments come from the companies involved themselves and are product statements, not an independent evaluation. → TechRadar

Synthszr Take: Huang calls the Surface Laptop Ultra a work of love, and that captures the entire division of roles: Redmond says what the computer should be able to do, Santa Clara delivers the component that makes it expensive. In doing so, Microsoft correctly identifies its own homework, which is to make Windows a layer that assigns rights to an agent, logs its actions comprehensibly, and defines a clear point at which a human countersigns. More hangs on this for Microsoft than on any graphics processor deal, because an operating system that you can entrust with an autonomously clicking agent cannot be easily replicated.

Anthropic is putting five percent of its shares into a 'Public Benefit Fund'

Anthropic plans to contribute five percent of its own company shares to a newly created “Public Benefit Fund.” This was reported by the Wall Street Journal, citing a company announcement on Sunday. According to Anthropic, the fund is intended to be used flexibly: it can make traditional donations or distribute funds directly to “specific population groups,” such as people whose jobs are displaced by artificial intelligence. The company had previously pledged $200 million to research and cushion labor market instability; the new initiative is said to go beyond that, according to its own statements. CEO Dario Amodei has repeatedly spoken publicly about the social and economic disruptions caused by rapidly advancing AI. → Wall Street Journal

Synthszr Take: The board itself decides whether the five percent ever becomes more, which means the corrective measure depends on the very entity it is supposed to correct. A commitment without a threshold and without a date is accounting with reservations. The $200 million was real money that is being disbursed; five percent of shares is a quota whose value increases on its own with each new financing round, without anyone having to add more.

Volkswagen partner Rivian is transforming into a technology supplier

In an interview at TechCrunch Disrupt, Rivian CEO RJ Scaringe outlined that his company intends to license its technology to “many different brands” in the future. The manufacturer has developed its own AI models for the in-vehicle assistant and is building an end-to-end stack for automated driving with Uber, which is intended to run in both production vehicles and future robotaxis. In addition, a proprietary 5-nm processor is being developed in collaboration with Arm and TSMC. Rivian has a $5.8 billion joint venture with the VW Group for electrical architecture and software, which Scaringe says he would like to expand to include artificial intelligence and autonomy. The startup Also has emerged from an internal micromobility program, and Scaringe also founded the industrial robotics company Mind Robotics, in which Rivian holds a 31 percent stake and which lists the automaker as a potential customer. → TechOrange 科技橘

Synthszr Take: The fact that VW is investing $5.8 billion in a joint venture with a much smaller manufacturer says enough about the new pecking order in the industry. Rivian is licensing electrical architecture, driving stack, and AI models to other brands—exactly the parts that still allow a car to differentiate itself today. The 5-nm chip with Arm and TSMC is the real leverage here: the computing platform in the vehicle determines which functions can be retrofitted via updates years later.

Aether AI is building a robot brain designed to think in terms of cause and effect

The startup Aether AI is working on control software for robots that models relationships as cause and effect rather than as statistical patterns. The approach is presented in a video from Interesting Engineering. The idea behind it is that the system should be able to predict what an action will cause in the physical world before the robot performs it. According to the company, Causal Inference is intended to improve the transfer to situations that were not present in the training data. In contrast, common robotics models are trained on large amounts of recorded demonstrations and primarily learn correlations between observation and movement. → TechOrange 科技橘報

Synthszr Take: A model trained on correlation can grasp a cup and still not be able to say why it tips over. Causality is the old unsettled score in AI research, and in robotics, it immediately becomes expensive because the physical world acknowledges every fallacy with broken pieces. The claim of thinking in cause and effect is easily made and hard to prove: the proof would only be clean if the machine succeeds in an environment for which it has not seen a single demonstration.

After the depleted web, AI companies want data from everyday office life

Business Insider reports that AI companies are specifically looking for data from everyday work life to train agents and robots. The report’s reasoning: the open web has been largely exhausted as a training source after years of being collected via Scraping. What is now in demand is material about how people do their jobs, i.e., processes in enterprise software rather than published texts. The report names companies working on agents for office work and robotics as the buyers. → TechOrange 科技報橘

Synthszr Take: The open web was free and legally contested; data from everyday work is expensive to buy and therefore exclusive. It contains what’s missing on the web: the silent intermediate steps and the correction two seconds before sending. In March, we were still talking about Tokenmaxxing, i.e., how much model usage a workforce consumes; now the direction is reversing, and the usage itself is becoming a raw material.

China’s economy grows by 4.5 percent, driven by exports for AI data centers

China’s economy grew by 4.5 percent in the third quarter, recovering from a three-year low, as reported by the BBC, citing official Chinese data. The figure is below the 5 percent from the January-to-March period and thus at the lower end of Beijing’s annual target of 4.5 to 5 percent. According to the report, the growth was primarily driven by exports: global demand for computing chips and other components for building AI data centers is supporting the industry. China’s trade surplus is on track to exceed last year’s record of $1.2 trillion. The domestic picture is different: real estate investment fell by more than 20 percent in the first nine months compared to the same period last year, and an end to the slump is not in sight, according to Oxford Economics analyst Sheana Yue. → BBC

Synthszr Take: Western data center billions are currently financing the growth of the very economy they tried to slow down with export controls. Beijing has linked its economy to an investment cycle over which it has no control: if the construction wave for AI infrastructure in the US collapses, China will lack domestic demand as a buffer, as real estate investment has fallen by over 20 percent in nine months. Conversely, the same dependence applies, just the other way around, because no data center can start up without the components from this supply chain.

New Stanford architecture generates text in conceptual layers instead of token by token

Researchers at Stanford University have introduced Next-Concept Modeling (NCM), an architecture that no longer predicts text token by token, but first lays out an entire response as a hierarchy of discrete codes and then elaborates them from coarse concepts to fine details. In experiments with models of comparable size, NCM generated text 10 to 31 times faster than strong Diffusion Language Model comparison models, with the same or better quality, according to the paper. In selected language model benchmarks, the method also processed four to five times fewer training tokens. The training occurs in two stages: a multi-scale autoencoder learns to compress text into hierarchical codes and reconstruct from them, then a transformer learns to predict the codes of each level. The efficiency gain comes from the generation mechanism: NCM requires one sequential step per hierarchy level instead of per output token, while the codes within a level are generated in parallel. → AlphaSignal

Synthszr Take: A factor of 10 to 31 with comparable quality is a matter of calculation, not research romanticism. The reason for this is unspectacular: one generation step per hierarchy level instead of per token, and the early levels are short, and therefore cheap. Add to that the four to five times lower training token load, which means the whole thing is also efficient on the training side and not just in operation. Caution is still advised, because the comparison was against diffusion models and not against the powerful autoregressive systems that are in production today.

TypeSafe unlocks new features with Jev 1.14.0

TypeSafe AI has released a new version of its model Jev and is simultaneously unlocking several experimental features that are only accessible via its own API headers. Under the label “Danger Zone” are image processing, a new type, and an unlocked context length. The company writes that, in its own assessment, these functions are too volatile to be recommended for production scenarios. According to the provider, image support was already present in the previous jev-1.13.0 and can be used at any point in the JSON interface. A new addition is the box type, which returns a single Bounding Box and is only available in jev-1.14.0. TypeSafe had previously deliberately capped the context length because longer contexts reduce reliability; this limit can now be lifted via a header. → TheSequence

Synthszr Take: An HTTP header as the dividing line between production code and a playground is the most pragmatic answer to a problem that every model lab currently has: features are ready faster than they are reliable. TypeSafe sells determinism, and any shaky feature that quietly slips into the standard path immediately damages this promise. The Danger Zone makes the risk a conscious line in the request and visibly shifts the responsibility to the developer, which I consider more honest than the usual practice of simply including beta-quality features.