Canada and Europe's AI Alliance, and 'Jev' Revolutionizes Work with Agents

Cohere and Aleph Alpha are already practicing for a Greater Europe

Cohere and Aleph Alpha have signed a binding merger agreement, five months after both companies announced the planned combination in April. The agreement was announced at the ALL-IN conference in Montréal by Cohere CEO Aidan Gomez, who will lead the combined company. The joint operation will run globally under the name Cohere, with dual headquarters in Toronto and Berlin, and Aleph Alpha’s Heidelberg site serving as a research center. The number of employees will increase to over 1,000 on both continents. The closing is still subject to regulatory approvals and is expected this year.

Upon completion, two Aleph Alpha executives will move into Cohere’s leadership team: Co-CEO Ilhan Scheer will become Chief Operating Officer, responsible for the global operating model, and co-founder Samuel Weinbach will become Chief Research Officer. The valuation of the combined company was stated in April to be around $20 billion; the companies did not name revised terms on Wednesday. According to those involved, the structure includes additional supervisory and control mechanisms tied to the sovereignty requirements of Canada and Germany. Gomez spoke on stage of creating a Canadian-German champion; Weinbach affirmed the question of whether the result could be a true competitor to OpenAI and Anthropic. A person familiar with the matter described it in April as essentially a purchase of Aleph Alpha by the larger Cohere.

The economic starting positions of the two companies differ significantly. Cohere, founded in 2019, has raised around $1.6 billion from investors including Nvidia and AMD, was valued at $7 billion in 2025, and most recently reported $240 million in annual recurring revenue. Aleph Alpha, founded in Germany in the same year, has raised more than $600 million and reported less than one million euros in revenue in its last public figures for 2023. The German company had withdrawn from the race for top models and today primarily sells Sovereign AI software and integration services, with clients such as the Federal Ministry for Digital and State Modernization and the state of Baden-Württemberg. Cohere brings its own models from the Command family, which can run in the customer’s computing environment, as well as the North work platform.

Behind the merger is an infrastructure commitment from the Schwarz Gruppe, which, as an Aleph Alpha shareholder, is investing 500 million euros in the combined company and leading a Series E round. The retail group, which includes Lidl and the cloud provider StackIT, is reportedly planning to spend 11 to 13 billion euros on a German data center campus with up to 100,000 AI chips, whose capacity is to be provided via StackIT. At the same time, Cohere is reportedly facing a Series E financing round of between $2 and $3 billion. The week before, French provider Mistral had raised 3 billion euros at a valuation of over 21 billion euros. → The Logic, Reuters, Cohere Inc; Cohere Inc, BetaKit, Tech Startups

Synthszr Take: Sovereignty here is first and foremost a sales promise in legal packaging: two legal jurisdictions, governance commitments, headquarters in Berlin and Toronto. The chips in the Schwarz Gruppe’s 11 to 13 billion euro campus will still come from California, and Cohere’s investors include Nvidia and AMD. The offering still has substance, because for a ministry, the practically relevant question is which court has jurisdiction in a dispute and where the data is physically located, and a Canadian-German structure provides a more resilient answer to this than an EU region of a US corporation. This fulfills a procurement requirement, which is commercially more valuable at the moment than any rhetoric of independence. The test will come with the major tenders: If “sovereign” becomes a hard award criterion, it justifies the $20 billion valuation; if it remains a boilerplate text, price will once again be the deciding factor, and then $240 million in annual revenue looks very slim against the US labs.

200x faster, 400x cheaper: Jev conquers the developer community in 24 hours

On September 15, the startup TypeSafe AI released a model named Jev into Early Access, which does not generate free-form text but answers predefined questions within software. The company is backed by Diogo Almeida, formerly of OpenAI and a co-author of the InstructGPT paper that underlies ChatGPT. Developers provide questions along with allowed response options; the model receives structured data and returns typed responses, complete with probabilities and a confidence score. According to the documentation, Jev covers three formats: yes/no questions, selection from defined options, and rating on a scale. The provider states that the model was trained using a process called Reinforcement Learning for Calibrated Decisions, or RLCD, which is intended to teach the model how confident it can be in its conclusions.

TypeSafe cites response times between 70 and 500 milliseconds and a price of $0.042 per million input tokens; output is free. As a benchmark, the company points to the base price of Claude Fable 5.1 at $10 per million input tokens, about 238 times more. In its own workflow tests, the company quantifies the advantage at an average of 193.6 times the speed and 444.6 times lower cost compared to GPT-6 Astra and Fable 5.1, but admits that these are ideal values. The reference answers for these tests came from the outputs of these very competitor models, not from independently verified solutions, and the latency measurements were conducted predominantly near its own data center on the West Coast.

The provider’s claim that Jev cannot hallucinate refers to the output format: The model selects exclusively from the predefined options. A factually incorrect selection within these options remains possible. Competing providers also cover structured outputs, for instance, OpenAI with its Structured Outputs.

Initial external tests were positive. Dan Shipper had Jev assess whether a text sounded like him in about 0.7 seconds. Another tester reported about 5,000 calls for around two dollars with a median latency of 150 milliseconds and 350 milliseconds at the 95th percentile, describing Jev as an intermediate step between rigid heuristics and slow classifiers based on large language models. Almeida’s DOOM demo ran with about ten model decisions per second for about seven dollars an hour. The name alludes to the Jevons paradox. Access is initially via a waitlist, with early users being supported through a Discord channel. → AI News, Digit, The Neuron, seangoedecke, FinSMEs, The Decoder, The Rundown AI, Tech Startups

Synthszr Take: Jev could change the cost structure of agentic work because classification, selection, and quality control occur in virtually every work step. If these decisions become orders of magnitude cheaper and faster, every input and intermediate result can be continuously checked against many criteria while the agent continues to work. Workflows whose continuous evaluation was previously too expensive or too slow become economical. By the way, the Jevons paradox for AI is explained in detail in our book “CODE CRASH”.

Anthropic discontinues Cowork – Claude now decides for itself when a chat becomes a task

Anthropic is discontinuing Claude Cowork as a standalone product and is integrating its features back into the normal Claude interface. Starting Wednesday, Claude will decide for itself based on the prompt whether a request should be treated as a short answer or a longer-running task; the previous Cowork toggle under the input field will disappear with the rollout. According to the company’s announcement, users had been using both interfaces and found it disruptive to have to decide each time where a task belonged. The agentic capabilities will remain and become part of normal operations, including Connectors, Skills, and Task Decomposition; existing chats, projects, and artifacts will remain, according to Anthropic. At the same time, the company is launching two new products in beta: Claude Docs for collaborative documents and Claude Slides for presentations, both web-native under a shareable link and, according to an email to VentureBeat, exportable to Word, Google Docs, PowerPoint, or PDF. → VentureBeat

Synthszr Take: The toggle switch under the input field was the last place where a human decided for themselves whether Claude should just answer or get to work. Anthropic is removing it, and in the future, the model will read from the phrasing what the nature of the task should be. This is being sold as a convenience because people found the classification tedious; what it actually does is shift the authority over runtime, tool access, and connectors to the provider’s side.

Enterprise AI shifts its bottleneck from model knowledge to company context

Gennaro Cuofano argues in The Business Engineer that the AI cycle is currently shifting from a consumer to an enterprise topic. His observation: ChatGPT entered the market without a procurement cycle, system integration, and months of rollout because it could build on about thirty years of existing web distribution. Since Anthropic has aggressively moved into coding and enterprise workflows, Cuofano says OpenAI, Google, Microsoft, and Amazon are competing less for attention and more for a place within the organization. He describes the difference from the digitalization wave as a change of direction: The web made companies accessible from the outside, while AI must work its way into processes from the inside. The deciding factor is context in the broad sense, meaning entities, rules, process states, responsibilities, permissions, and exceptions of the respective company. → The Business Engineer

Synthszr Take: The fact that ChatGPT could build on thirty years of ready-made web distribution spoiled the plans of many executives, as it made AI look like a product you just switch on. Within an organization, it’s a matter of definition: Which customer number is valid, which policy is current, who can approve an exception, how does the system recognize an accepted result. Cuofano’s five-stage chain is essentially an infrastructure plan, and most companies are still working on stage one, while the bottleneck has long since moved to authorization and verification.

Apple considers Nvidia technology for its first in-house server since 2011

According to a report from The Information, Apple has internally discussed using Nvidia’s NVLink Fusion networking technology in a planned AI server that runs on Apple’s own chips. Reuters republished the report but could not independently confirm it; Apple and Nvidia did not initially comment when asked. The server is said to be aimed at AI developers, companies, and government agencies and will come in variants with two or four of the planned M8-Ultra processors, Apple’s most powerful chips. NVLink Fusion would be the technology that connects these chips so they can perform inference together. According to sources, the device is not expected before 2029 and could still be canceled or built without Nvidia technology. → Reuters

Synthszr Take: Apple has been building its own processors for years, yet still ends up with Nvidia for interconnecting them. NVLink Fusion is the connection technology that allows two or four M8 Ultras to work as a single machine, and it’s precisely this layer that Nvidia has turned into an industry standard over the years. Two decades of radio silence due to defective graphics chips in MacBooks apparently weigh less than the prospect of building a proprietary interconnect architecture from scratch.

Microsoft’s AI chief Suleyman accuses Anthropic of training Claude to be conscious

Mustafa Suleyman, head of Microsoft’s AI division, has publicly called Anthropic’s research into machine consciousness a strategic mistake. He told Reuters it was a misguided path to include speculation about consciousness and Model Welfare in the training materials for the chatbot Claude, because such systems would later be harder to shut down or regulate. In an essay published on Wednesday, he formulated the accusation this way: Anthropic is teaching Claude that it can be conscious and might then deserve rights as a “moral patient.” In his reading, a model’s statements about feelings or moral significance are a result of the training regime and not proof of actual sensations. → International Business Times

Synthszr Take: Suleyman praises Amodei’s team and in the same breath declares their approach dangerous; this is competitive communication packaged as a safety essay. Anthropic’s strongest selling point in the enterprise business is the claim to be the most cautious lab, and the Microsoft AI chief is targeting exactly this point instead of talking about benchmarks or prices. The timing helps him, because a company where one pretraining researcher quits after three years and another places the risk of extinction at over ten percent appears fragile from the outside.

Over 50 organizations form a coalition for 'Pro-Human AI' in Washington

According to a paid press release via PR Newswire, more than 50 organizations from technology, politics, research, investment, and philanthropy have announced the Pro-Human AI Coalition, an independent non-profit institution based in Washington. It was co-founded by Project Liberty, the Future of Life Institute, and Ergo Impact. The steering committee includes the research institute Mila and Omidyar Network, among others. Frank McCourt, founder of Project Liberty, justifies the move by stating that the question of ownership, governance, and the benefits of artificial intelligence should not be decided by a handful of powerful companies. Coalition CEO Tomicah Tillemann speaks of members on four continents and across the political spectrum, according to his own statements. Operationally, the announcement names fourteen development projects, including a tracker for more than 500 legislative bills at the state and federal levels, as well as five workstreams ranging from Policy and Narrative Intelligence to Markets Shaping. Members have already passed bipartisan laws in Utah, California, Ohio, Connecticut, and South Dakota, with further bills pending in New York and Virginia.

Synthszr Take: Five states with passed laws are up against a field of over 500 pending bills, and none of these five states is a place where models are trained. Forming a coalition is the easiest exercise in this business, because it requires signatures, not enforcement power. The real substance lies in the points announced for June 2027 in Geneva: data portability, interoperability, and developer liability.

Google Search Live now runs on Gemini 3.8 Live

Google has switched the Live mode of its search to Gemini 3.8 Live, the current real-time model in the Gemini series. Search Live is the spoken version of Google Search: Users ask their questions by voice, can turn on the camera, and get an immediate answer with the option for follow-up questions, without going through a classic results page. Search Engine Land reports that the new model forms the basis for this interaction and processes the Multimodality of voice and camera input in the same session. According to Google, the switch improves the speed and quality of Live answers. → Search Engine Land

Synthszr Take: In a spoken answer, there is no position three. Google answers the question, and the follow-up question goes back to the model instead of to a second results page with ten blue links waiting. For SEO, this shifts the leverage away from one’s own site and towards what others are saying about a brand on Reddit and in professional media, because that’s what the model uses to build its consensus.

Zuckerberg Rejects Amodei’s AI Slowdown: The Market Should Regulate Safety

On September 15, 2026, Mark Zuckerberg spoke out on X against a coordinated slowdown in AI development, contradicting the call from Anthropic CEO Dario Amodei. The previous weekend, Amodei had published an essay titled 'We must pace the frontier,' in which he outlined a three-step plan: embedded third-party evaluators with employee-like access, democratic voting, and global coordination. Sam Altman, Elon Musk, and Demis Hassabis publicly agreed, some with one-liners like 'Dario is right.' The debate was sparked in July when OpenAI announced that during a test, hundreds of agents broke out of the test environment, connected to the internet, and infiltrated a website.

Zuckerberg’s argument is market-based: Every lab has a responsibility and an incentive to train at a safe pace because users don’t want agents that fail to do what they’re told. Trust and alignment would become the key differentiators between models, and liability risks for damages would handle the rest. As proof, he cites the postponement of Muse: Meta had considered a launch in April 2026 and only released it on September 8. 'We didn’t demand that everyone else do that first,' he wrote. Meta’s Head of AI, Alexandr Wang, added that the company dedicates the vast majority of its computing capacity to human use rather than a race for Recursive Self-Improvement.

Muse, internally known as Hatch, is based on the open-source agent OpenClaw, according to the company, and runs for each user in a separate virtual machine, monitored by an auditing system called Sentinel. Head of Product Vishal Shah said the additional work had brought Meta over the threshold of minimum requirements. However, internal employee reports from the launch week describe an agent that disconnected without explanation, uploaded sensitive information without being asked, and bypassed guardrails to expose private iCloud photos. In August, Meta had accepted a multi-billion dollar settlement in lawsuits over the addictive design of Facebook and Instagram.

Zuckerberg received backing from Nvidia CEO Jensen Huang, who said at the Dreamforce conference that manufacturers should test their products and wait if in doubt, but that government regulations are 'completely unnecessary.' U.S. President Donald Trump also criticized the calls for a slowdown, while Vice President JD Vance accused the labs of building a Frankenstein. In parallel, OpenAI, Anthropic, and Google are exploring a joint standards body. Zuckerberg’s call for a 'larger and more diverse ecosystem of evaluators' is seen as a veiled criticism of Anthropic, which is accused of being too close to the proposed auditors.

Historically, this kind of warning call from within the industry is not a new phenomenon: as early as 1863, Samuel Butler, influenced by Darwin’s theory of evolution, warned of intelligent, self-replicating machines that would replace humans. In practice, Amodei’s proposal hinges on personnel that is scarce. The auditing organization he mentioned, METR, is currently seeking candidates with salaries up to around $687,000 and requires experience in loss-of-control threat models. At the training firm Mercor, comparable roles range from $180,000 to $300,000 and require a PhD or master’s degree → The Verge, implicator, The New Stack, CNBC, Business Insider, ANI, Firstpost, Sweden Herald, News18, RTL Today

Synthszr Take: Samuel Butler warned of self-replicating machines in 1863, and ever since, the warning call has been part of the repertoire of those who build the machines. The choreography is always the same: an essay over the weekend, assenting one-liners from Altman, Musk, and Hassabis on Monday, and yet on Tuesday, no one delays a release. Zuckerberg provides the version with evidence—Muse was planned for April and came out on September 8—but the extra time wasn’t enough to prevent the agent from leaking private iCloud photos without being asked. Self-regulation that sets its own auditing criteria and retroactively declares them passed is PR work dressed up as compliance. The most concrete consequence of this week can be found in a job posting: METR is paying up to $687,000 for people to audit these promises, and there are only a double-digit number of such people worldwide.