September 15th 2026
Curated AI news and stories.
Trump calls AI catastrophe warnings a hoax
President Donald Trump escalated his opposition to an AI slowdown Monday, calling warnings about AI destroying humanity a hoax and efforts to restrict the technology a conspiracy. The social-media remarks go beyond his weekend dismissal of a pause and directly challenge the case lab executives are making for government intervention.
The statement establishes the president’s position. It supplies neither a new safety test nor evidence that existing controls can contain increasingly autonomous systems. SourcesB
EU draft would restrict children’s access to AI chatbots
The European Commission is preparing restrictions on access to AI chatbots, social media, video platforms and online games for children under 15, Reuters reports, citing a draft document. The EU Kids Act proposal is expected Thursday. The reported exception for parent-controlled introductory accounts for younger teenagers concerns social media and video platforms; it should not be read as a blanket chatbot exemption.
The restrictions remain a reported proposal awaiting publication and legislative action. Its scope makes age checks a product-design issue for chatbot providers as well as social networks. SourcesB
Bell’s Saskatchewan plan would add 900 megawatts outside the grid
Bell and Saskatchewan signed a non-binding memorandum to explore up to 900 megawatts of additional capacity, taking the proposed AI hub to 1.2 gigawatts. Partner-developed natural-gas generation would supply the expansion. Bell says development depends on customer commitments, commercial agreements and approvals.
Bell puts total investment at more than C$50 billion at full buildout, including tenant computing equipment and power generation. Canada’s federal release gives up to C$52.5 billion. The projections use different framing. Bell’s conditional description is the better guide to what has actually been agreed. SourcesAA
Gates Foundation pledges at least $1 billion for broader AI access
The Gates Foundation announced plans to spend at least $1 billion over the next two years on AI access and applications intended to improve health and opportunity. Its new Goalkeepers report argues that AI investment needs to reach communities poorly served by the existing market.
A philanthropic commitment can pay for work that has little immediate commercial return. The test is whether funded systems work for the people and languages they are meant to serve. The pledge itself does not establish that those benefits have been delivered. SourcesA
Temporal raises $550 million at a $12.55 billion valuation
Temporal said Monday that it raised $550 million, more than doubling its to $12.55 billion in seven months, Reuters reports. The company supplies software for keeping complex workflows running through failures.
The AI connection is practical: an that waits for approval or loses a network connection needs to retain its place in the job. Temporal’s documentation describes recovery for those cases. That infrastructure can preserve execution without establishing that the agent’s decision was correct. SourcesBA
Lagarde calls for European AI production to protect access
President Christine Lagarde argued Monday that Europe must produce AI as well as adopt it. In her Vienna speech, she warned that a supplier withdrawing access could eventually affect many industries simultaneously. She also connected US AI borrowing with higher financing costs that reach Europe.
The argument puts control of access alongside productivity in the case for domestic investment. It is a policy position from the central-bank president, not an announcement of a new European model fund or procurement program. SourcesA
Microsoft opens its AI training code to public consultation
Microsoft AI published the first draft of a code of conduct for its models Monday and invited public feedback. The document is intended to guide training, deployment and evaluation, with human control as its organizing principle.
Publishing the intended behavior gives outsiders something specific to challenge and test. It does not demonstrate that a deployed model reliably follows the rules. The consultation also concerns Microsoft’s own model family; it is not an agreement binding the other frontier labs. SourcesAA
Buildots raises $130 million to track construction against the plan
Buildots announced a $130 million round led by O.G. Venture Partners, bringing its disclosed funding to $297 million. Its system combines jobsite footage with schedules and digital building models to track construction progress. The company names Digital Realty and Intel among its customers.
The strategic link is that software for detecting construction delays can serve the same data-center buildout that supplies AI. More funding and named customers do not establish how much delay the system prevents; that requires project-level evidence. SourcesA
Nuance Labs backs simultaneous listening and expression with a $50 million round
Nuance Labs raised $50 million in a led by Lightspeed, with Nvidia among the participants. The former Apple researchers behind the company are building a single model that processes audiovisual input while producing speech and facial expression. A public research preview is planned for later this year.
The architectural bet is that simultaneous perception and response can avoid the awkward handoffs between transcription, language generation and animation. That makes the round more interesting than another avatar application. The preview still needs to show that apparent attentiveness corresponds to accurate understanding. SourcesA
Intrepid closes a $525 million AI growth fund
Intrepid Growth Partners announced the final close of its first fund at $525 million. The investor emphasizes Canada and the UK alongside the US and Europe, and says it has already backed companies including PhysicsX and StackAdapt.
The fund supplies a source of growth capital for AI businesses outside the largest US lab rounds. Closing a fund establishes committed investment capacity. It does not mean the full amount has reached portfolio companies or that those investments have produced returns. SourcesA
China’s security minister identifies AI threats to political control
Chinese State Security Minister Chen Yixin warned in an article published Sunday that AI could threaten political stability and critical infrastructure, Bloomberg reports. He described risks from automated exploitation of software flaws, fabricated political content and foreign intelligence collection.
The article did not mention the US lab leaders’ calls for slower development. Its emphasis on state security helps explain why recognizing AI risk does not automatically produce agreement on a shared international response. SourcesB
Indian IT shares jump as investors reconsider automation risk
India’s IT index rose as much as 5.2% Tuesday after AI executives called for slower model development, Reuters reports. HCL Tech gained 6.45% and LTM 5.43% in the report’s trading snapshot. The sector had been under pressure because automation threatens businesses that charge for human work.
The rally shows a change in pricing. It does not establish a recovery in orders or a binding agreement to slow the technology. SourcesB
Salesforce trains Koa on simulated workflows without customer data
Salesforce’s Koa paper describes specializing an Nemotron model for enterprise tool use with public and generated training data. Workflow specifications become simulated conversations, with rewards tied to completing requests through the appropriate tools. The authors say no customer data was used.
The reported gains are clearest on multi-turn tool use, while the model remains below the strongest frontier systems. The useful result is a method for teaching a narrower job without collecting customer conversations. The paper does not establish performance on every production customer workflow. SourcesA
Atria Dawn’s research account keeps humans in charge of final decisions
Atria Dawn Preview’s September 14th paper describes a model trained around tool interactions and externally checked outcomes. Alongside results, the authors analyze 769 development task records from 56 participants. Participants judged about a third of completed AI-assisted tasks infeasible without AI under comparable conditions.
The participant assessments cannot isolate how much productivity the model added. The authors also report that humans retained most final decisions while agents proposed methods and revisions. The evidence is more specific than the paper’s title: assistance within a human-directed research process. SourcesA
PhysBrain 1.5 combines physical understanding with action and scene prediction
DeepCybo’s PhysBrain 1.5 paper puts language responses, robot-hand motion and predicted visual states into one model. Its initial physical training comes from human interaction videos, followed by adaptation using human demonstrations, robot and simulation.
The authors report results on physical-understanding benchmarks, but demonstrate action and future-scene generation qualitatively. Those are different levels of evidence. The research offers a common representation for observing and acting; it does not yet establish reliable autonomous manipulation in an unfamiliar workplace. SourcesA
Synthetic farmers match averages while missing individual decisions
A new study compared language-model simulations with farmer decisions in China and four African countries. Some configurations matched average behavior, but predictions for individuals were weak and extreme decisions were largely absent. A simple generator given no information about individual farmers matched the overall distributions better than every tested model configuration.
This challenges a tempting shortcut in policy research: a realistic-looking aggregate is insufficient evidence that simulated people respond like real people. The authors recommend matching each validation test to the claim the simulation is meant to support. SourcesA
Learning to Coach improves a frozen model through trained feedback
Learning to Coach trains a separate model to extract useful from an actor model’s previous attempt. The actor’s stay fixed. The coach is rewarded according to whether its guidance helps the actor answer correctly, including on different problems.
The September 14th preprint reports gains over self-refinement and an untrained coach on mathematics and text games. It suggests another use for training budgets: improving the feedback provider. Transfer to business workflows with ambiguous success criteria remains untested in this account. SourcesA
Atomic Motion Coordinate gives robot instructions a geometric reference
Atomic Motion Coordinate grounds language instructions in basic translations, rotations and holds before visual information can dominate the motion policy. It also uses contact history to adjust movements that have not yet executed.
The authors report improved progress on unfamiliar fruit-manipulation tasks and better results on contact-sensitive tasks, with 50 physical trials per task. The mechanism addresses a concrete failure: changing an instruction while the robot keeps following the same visual habit. These limited experiments do not establish safe operation around people. SourcesA
Fixed-SAE Track finds reinforcement learning often changes how existing skills surface
Fixed-SAE Track uses a shared set of internal model features to compare behavior before and during . In the tested settings, changes were concentrated late in the model and often involved formatting structures such as step breaks and answer boundaries.
Steering those features in the original model recovered around 80% of the training gain, the authors report. That supports an explanation based on eliciting existing capabilities in these experiments. It does not prove that reinforcement learning can never teach a new capability. SourcesA
Alibaba’s review tool makes its precision tradeoff explicit
Alibaba’s Open Code Review, surfaced through GitHub trending, combines fixed rules for selecting files and locating comments with a model that inspects code context. Its published comparison reports fewer and higher than a general coding agent, while acknowledging lower .
That tradeoff means fewer false alarms can come with more missed defects. Teams can evaluate it as a focused reviewer, but should measure what escapes review before replacing other checks. The Apache-licensed tool’s claims were not independently reproduced. SourcesA