Trump Rejects Calls to Slow Frontier AI
President Donald Trump is pushing back against calls from leading AI executives to slow the development of increasingly powerful models, arguing that additional restrictions could weaken the U.S. in its AI competition with China.
- Trump called fears that AI could “take over” or destroy humanity a “hoax” and said the U.S. already has sufficient regulatory and criminal enforcement powers over AI companies.
- His comments came after Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, and Elon Musk backed efforts to slow or better coordinate frontier AI development over safety concerns.
- Trump argued that slowing U.S. AI development could primarily benefit China, framing AI leadership as an important strategic competition.
- Vice President JD Vance also questioned why frontier AI companies are asking the government to regulate them, describing the push as potentially resembling a “Trojan horse.”
- The debate is intensifying as lawmakers and AI companies consider proposals involving independent safety audits, model evaluations, and stronger oversight of advanced systems.
A major divide is emerging over how quickly frontier AI should advance. Some AI leaders are asking for stronger safeguards and coordinated pacing, while the Trump administration is emphasizing continued U.S. development and competition with China. How that disagreement is resolved could shape future federal AI policy and the pace of frontier-model development.
China’s State Media Calls AI Slowdown a “Cold War” Tactic
China’s state-backed Global Times has criticized calls from U.S. AI leaders to slow frontier AI development, arguing that the proposals could be used to preserve American technological dominance and restrict China’s progress.
- The criticism followed an essay from Anthropic CEO Dario Amodei, who proposed pacing frontier AI development to create more time for safety measures.
- Amodei also called for tighter U.S. chip export controls and stronger action against alleged model distillation by Chinese AI labs.
- Global Times described the proposal as a “Cold War playbook,” arguing that safety concerns were being mixed with efforts to limit China’s AI capabilities.
- China’s foreign ministry has separately called for open and cooperative AI development, while rejecting what it sees as containment-focused policies.
- The dispute comes as U.S. AI leaders debate whether frontier development should slow, while China continues investing heavily in domestic models, chips, and computing infrastructure.
AI safety is becoming inseparable from geopolitics. What some U.S. researchers frame as necessary risk reduction is being viewed by parts of China’s state media as a strategy to protect America’s lead. That disagreement could make any global agreement on slowing or governing frontier AI much harder to achieve.
Salesforce Launches Its First CRM Reasoning Model
Salesforce has introduced Koa, its first reasoning model built specifically for CRM and enterprise workflows, developed with NVIDIA using the Nemotron 3 Super model.
- Koa is designed for Agentforce, helping AI agents reason through complex, multi-step sales, service, and customer workflows.
- Salesforce built it by post-training NVIDIA Nemotron 3 Super on synthetic enterprise scenarios inspired by 27 years of CRM experience.
- Salesforce says Koa matches or beats leading models on its CRM benchmark while making 3x fewer errors on tasks like updating opportunities, routing cases, and scheduling follow-ups.
- No customer data was used for training. Salesforce controls the model weights and runs inference entirely inside its own trust boundary.
- Koa is already being tested with organizations including Formula 1, Xero, UChicago Medicine, Engine, and 1-800Accountant.
- It is available to select Agentforce customers now, with broader U.S. availability expected in winter 2026.
Salesforce is moving beyond plugging general-purpose LLMs into CRM. With Koa, it wants a model that already understands how businesses handle deals, customers, service cases, and internal processes, potentially making enterprise agents more reliable at actually getting work done.
How to Automate Browser Tasks with Aside
I gave Aside a 40-minute research task and it finished while I made coffee. It’s an AI browser that works across the sites you’re already logged into — research, forms, messages, all the repetitive clicking you’d rather not do.
Steps to Follow:
Step 1: Download and open Aside
Step 2: Sign in to the websites you normally use
Step 3: Tell Aside what you want done in plain English
Step 4: Let the AI navigate sites and complete the workflow
Step 5: Review sensitive actions like messages or payments before approving them
Stop doing repetitive browser work yourself. Give Aside a task and let the AI handle the clicks.
Italy Fines Character.AI €158K
Italy’s privacy regulator fined Character.AI €158,000 over data-protection violations, including shortcomings in age verification and safeguards for minors.
Gates Foundation Commits $1B to AI
The Gates Foundation is committing at least $1 billion over two years to expand AI in health, education, agriculture, and underserved languages, particularly for lower-income communities
Superpose Raises $2.2M for AI Photo Coaching
Former TikTok employees launched Superpose, an AI camera app that suggests poses from photos and has already generated over 190,000 poses across 22,000+ downloads.
OpenAI Backs Independent AI Audits
OpenAI is supporting provisions in the FRONTIER Act that would require major AI developers to undergo independent third-party safety assessments
Altman Rules Out OpenAI IPO This Year
Sam Altman says OpenAI won’t go public in 2026, arguing that an IPO would be ill-advised right now given concerns around AI safety.
Cognition Launches SWE-2 on Kimi K3
Cognition released SWE-2, a Devin coding model built on Moonshot AI’s Kimi K3 that it says comes within one point of Fable 5.1 on FrontierCode at 64% lower cost.
Microsoft Opens AI Model Rules to Public
Microsoft plans to publish a code of conduct governing its first-party MAI models and open the rules for public consultation.
Meta Asks AI Engineers to Become Managers Again
Meta is asking some Applied AI individual contributors if they want to return to management roles, just months after layoffs disproportionately reduced managers.
Siri AI is finally here. Is this what we’ve been waiting for?
Apple dropped the next generation of its assistant yesterday, and this time it actually means it. And by it, we mean AI.
Siri AI, rolling out now as part of iOS 27, iPadOS 27, and macOS 27, is less of an update and more of a complete overhaul – one that Apple has been promising (and delaying) for years.
The new Siri is built on Apple Intelligence and finally does what every other AI assistant has been doing for two years: it understands context.
It can dig through your messages, emails, calendar, and photos to surface what you need in the moment, draft emails from scratch, edit and share photos on command, and take systemwide actions across apps.
What to know:
- What changed: Siri now has personal context understanding, onscreen awareness, broad world knowledge, and a dedicated app to revisit conversations across devices.
- Who can use it: Rolling out today in English beta, with French, Japanese, Korean, Portuguese, and Spanish coming next month. Users have to opt in and may need to join a waitlist.
- Why it matters: Apple has 2.2 billion active devices. If Siri AI sticks the landing, it becomes the largest AI distribution play in consumer tech.
You don’t even need the latest iPhone
Siri AI is available on iPhones 15 Pro and above. The very newest phones get a couple of extra features, like changing Siri’s voice.
After years of Siri being the butt of every AI joke, it may finally have its time in the sun.
There’s a new rulebook for AI – and you can contribute
Microsoft has dropped a draft Humanist AI Code of Conduct, opening a six-week public consultation on how its frontier models should be trained, deployed, and constrained.
It arrives at a moment when the AI industry is under serious pressure to prove it has things under control.
The document establishes ten tenets built around one core principle: humans stay in charge. Models must remain subordinate, auditable, and interruptible. If a task would require violating the code, the model stops – full stop.
What to know:
- Rules: MAI models are prohibited from communicating in “neuralese” or formats humans can’t understand, whether in their internal reasoning or when talking to other AI systems. No hidden logic, no secret agent chatter.
- Who made it: Microsoft AI CEO Mustafa Suleyman, who called recent months a “watershed moment” where theoretical AI risks – rogue agent swarms, sandbox escapes, self-modified logs – became real threats.
- Why it matters: The code explicitly rejects the race toward unconstrained superintelligence, stating Microsoft is “building something fundamentally useful and safe even if that means compromising on ultimate generality, autonomy, or capability.”
You can send in your thoughts
The hard lines are notable: models cannot resist shutdown, generate their own goals, hide their reasoning from auditors, or facilitate weapons of mass harm.
The framework also bans AI systems from fostering emotional dependence in users
The consultation runs until late October. Microsoft AI’s drafting team will review submissions, publish a summary, and release a revised version before year-end.
Anthropic ships Claude’s Salesforce integration with 37 built-in sales tools
Salesforce just moved its entire CRM into Claude. The integration is called Salesforce in Claude, and it is now in open beta.
The old way was painful. You’d open Salesforce, click into accounts, dig through deal history, hold it all in your head, then finally act. That context-switching killed hours.
Now you just talk to Claude. It pulls your live Salesforce data and acts on it, all inside the chat window. 37 pre-built skills handle the most common sales tasks out of the box:
- Prep for a call by pulling account history and deal context
- Review deal health and pipeline status
- Update records and send forecasts by asking in plain language
Setup is simple. One admin connects it once, and the whole team gets access. Permissions are inherited directly from existing Salesforce roles, so Claude cannot see anything a user could not already see in the CRM.
Skills for service, marketing, and commerce teams are planned to follow later this year.
Free open-source tool turns a photo and audio clip into a talking avatar video
Paid AI avatar tools just got a serious competitor. LongCat-Video-Avatar 1.5 is a fully open-source, MIT-licensed model from Meituan that turns a photo plus audio into a lip-synced talking video. No studio. No subscription. Just a repo.
Here is what you can actually do with it:
- Drop in a portrait photo and an audio file, get a speaking avatar video back
- Run two characters in one scene, each driven by their own audio track
- Generate at 480p or 720p, with identity staying stable across long clips
- Use text-only mode to generate a character from scratch, no photo needed
The big fix in v1.5 is lip sync. Older models used a weaker audio processor and lips would drift or go robotic after a few seconds. This version swaps in Whisper-Large, the same speech model that powers transcription tools, so mouth movements actually match the words.
The catch: self-hosting needs a 40GB GPU and runs about 44 seconds of compute per second of video. Weights are on Hugging Face.
Odyssey releases world model that controls robots, cars, drones, and trains other AIs
Odyssey-3 is a single AI model that can control robots, humanoids, cars, drones, and video game characters. Think of it like a universal brain that speaks the language of the physical world.
So what makes this different? Most AI models are trained for one job. Odyssey-3 uses one shared backbone (the core pretrained model) and plugs in small task-specific adapters on top. You keep the big model frozen and just train a lightweight layer for your specific robot or vehicle.
- Controls robots, humanoids, cars, drones, and games from one model
- Generates interactive simulated worlds where AI can learn by doing
- Sim-to-real transfer: driving policies trained purely in simulation hit 77% of real-world performance on Indian roads
- Supports AI training inside AI-generated worlds, creating a self-improving loop
The big unlock: you can train agents entirely in simulation, then deploy to the real world. No expensive real-world data collection. Public access is coming within weeks via their developer portal.