Prez Reinforces AI Priority; New AI a 'Decider'; China and RSI

Today's AI Outlook: 🌥️

Trump Talks Guardrails, and Speakerphones

President Donald Trump joined an AI safety debate at the All-In Summit in Los Angeles by calling Nvidia CEO Jensen Huang, who put him on speakerphone for the audience.

Trump dismissed fears of an AI takeover and argued that slowing U.S. development would benefit China.

In a separate social media post, beyond what was reported in yesterday's newsletter, he described a strong, smart president as the guardrail the technology needs. His remarks signaled opposition to calls for a coordinated slowdown in AI development.

Why It Matters

The comments put presidential support behind continued AI development as parts of the industry call for greater caution. Nvidia supplies the computing chips used to train and run many AI systems, giving it a direct commercial stake in demand for that infrastructure.

The Deets

• Trump joined the Los Angeles conference by phone while Jensen Huang was on stage.
• He characterized fears of AI taking over as a hoax and framed continued U.S. development in terms of competition with China.
• His social media remarks emphasized presidential leadership as the safeguard for AI, rather than detailing a regulatory framework.

Key Takeaway

Trump publicly rejected arguments for slowing AI development. The remarks explain his position but do not establish what specific safeguards will be adopted.

🧩 Jargon Buster - Guardrails: Rules, technical protections and oversight intended to reduce the risks of developing or using AI.

The Token Bill Has a Favorite

Cheaper models are capturing a large share of activity on OpenRouter, the service that lets developers access models from different providers. Its Sept. 14 rankings, as reported today, placed seven Chinese models in the top 10, accounting for about 72% of tokens in that group. DeepSeek and GLM also featured prominently in agent workloads, suggesting that the cost of repeatedly running AI is shaping which models developers choose.

Why It Matters

For agents that call a model repeatedly, a small price difference can become a large operating expense. These rankings offer a useful snapshot of buying behavior, though they do not measure the entire AI market.

The Deets

• Chinese models accounted for roughly 72% of tokens among OpenRouter’s top 10, according to the reported Sept. 14 rankings.
• DeepSeek and GLM together represented about 62% of OpenClaw’s top-10 model tokens in the cited breakdown.
• GLM 5.3 Flash usage in the cited Claude Code breakdown was 4.4 times that of Claude Opus 5. These figures should not be read as total worldwide usage of either product.

Key Takeaway

Model selection increasingly demands a cost-and-quality test on the actual work. A famous name does not settle the invoice.

🧩 Jargon Buster - Tokens: The small pieces of text or other data a model processes, commonly used to measure usage and calculate charges.

🛠️ Tools & Products

New AI That Makes Fast, Calibrated Decisions

TypeSafe, founded by former OpenAI researcher Diogo Almeida, has emerged from stealth with Jev, an AI system that answers preset questions inside software instead of generating prose. It chooses among options defined by developers and returns confidence scores, with potential uses including sorting requests, scoring records and checking another AI’s output for jailbreak attempts. The company says this narrowing makes Jev dramatically faster and cheaper to run.

Why It Matters

Many software tasks need a reliable choice among known options. A system designed for that narrow job could make automated decisions practical in places where a general chatbot is too slow or expensive.

The Deets

• TypeSafe quotes $42 per billion input tokens, with no output charge.
• The company reports response times of 70 to 500 milliseconds and estimates a price 238 times below Claude Fable 5.1.
• TypeSafe says Jev cannot hallucinate because its answers are constrained to predefined options. That design does not establish that every selected answer is correct.

Key Takeaway

Jev’s appeal is inexpensive, fast decision-making inside software. Its accuracy still needs testing on the decisions a business actually makes.

🧩 Jargon Buster - Confidence Score: A model’s estimate of how certain it is about an answer, which needs calibration before people can treat it as a reliable probability.

Salesforce Brings the Brain In-House

Salesforce has introduced Koa, an in-house reasoning model for sales and customer support agents, built on Nvidia’s open Nemotron 3 Super model. The company trained it on synthetic business scenarios across more than a dozen industries and hosts it within its own systems, so customer requests do not have to go to an outside model provider. Its pitch is a model tailored to everyday CRM work, including updating deals and routing support tickets.

Why It Matters

A business-specific model gives Salesforce more control over operating costs, data handling and agent behavior. The quality question is whether those gains hold up in customers’ real workflows.

The Deets

• Salesforce says training used fully synthetic data, including simulated difficult support calls and sales conversations, without customer data.
• On an internal CRM benchmark, the company reported three times fewer errors than leading models. That is a company test, not independent validation.
• Salesforce also introduced AIforce to make CRM data and permissions available to outside AI tools, and opened ClaudeForce to all customers in beta.

Key Takeaway

Salesforce is taking direct responsibility for more of its agents’ reasoning while keeping connections to outside AI tools.

🧩 Jargon Buster - Synthetic Data: Artificially created examples used to train or test a model, such as simulated customer conversations.

🔬 Research & Models

China's Roadmap to RSI

More than 30 researchers, including contributors from ByteDance, Tsinghua University and Shanghai AI Lab, have published a roadmap for recursive self-improvement: AI systems helping design and improve their successors.

Their paper, “The Last AI Built by Humans,” proposes five levels of increasing autonomy and classifies 491 existing papers against that framework. Most of the research falls in the first two levels, where humans still shape much of the improvement process.

Why It Matters

A shared framework can help separate modest automation of research tasks from far-reaching claims about AI designing itself. The paper offers a way to describe progress; it does not demonstrate the final destination.

The Deets

• Level 1 executes human-designed upgrades; Level 2 diagnoses weaknesses and chooses fixes.
• Levels 3 and 4 expand control over what a model learns next and how it adapts after deployment; Level 5 redesigns the improvement process itself.
• About 75% of the 491 papers landed at Levels 1 or 2, and fewer than 6% at Level 5 under the authors’ classification.
• The authors see coding as a promising area because changes can be tested quickly. Robotics and scientific work have slower, more expensive feedback.

Key Takeaway

Watch for demonstrated improvements under clear tests. A five-rung ladder is useful, but drawing the top rung does not mean anyone has reached it.

🧩 Jargon Buster - Recursive Self-Improvement: A process in which an AI system helps improve itself, potentially making it better at carrying out further improvements.

⚡ Quick Hits

• Superhuman acquired Fathom, the AI meeting notetaker, to bring meeting context into its productivity and agent workflows. The reporting did not provide a purchase price.
• Legora is building a verified legal database to pair its AI tools with trusted research, putting it into more direct competition with LexisNexis and Westlaw.
• Odyssey introduced Odyssey 3, a world model the company says can control robot arms, humanoids, self-driving cars, drones and games. Release is expected in the coming weeks.
• OpenAI’s “Project Lily” drew scrutiny in a 404 Media report describing hundreds of contractors reading and rating real ChatGPT conversations, often without users’ knowledge.

🧰 Tools of the Day

• Gemini 3.8 Live: Google’s new voice models can continue speaking while they think. The reporting says the Extended Thinking version topped Artificial Analysis’ speech-to-speech quality ranking.
• StepAudio 3: A five-model audio suite covering voice agents, transcription and music. It is a relevant tool to explore for projects that need several kinds of audio processing.
• OpenRouter: Today’s practical guide uses the service to compare models and test less expensive image options. Its ad-mockup example produced useful directions but did not match the real product or the more expensive model’s version, a useful reminder to compare fidelity along with price.

Today’s Sources: The Internet, AI Secret, The Rundown AI