AI & design briefing for Wednesday, September 16, 2026

tldr

• Salesforce moved its CRM into Claude with 37 pre-built sales skills and announced Koa, its own domain reasoning model built by post-training Nvidia's Nemotron 3 Super. Marc Benioff's framing at the keynote: the SaaS-pocalypse was not the end of software, it may be the end of software that makes humans do all the work.
• Anthropic, Google, and OpenAI have discussed a third-party testing body, OpenAI backs the FRONTIER act's embedded evaluators, and Miranda Bogen at the Center for Democracy and Technology says outside testers cannot change company behavior without standards, mandatory fixes, and assigned responsibility when a model fails.
• Odyssey-3 controls robots, humanoids, cars, drones, and game characters from one shared backbone with small task adapters on top, and reports 77% of real-world driving performance on Indian roads for policies trained only in simulation.

Top Stories

• Salesforce moved its CRM into Claude and built its own model for the rest. At Dreamforce on Tuesday, Salesforce announced Koa, a domain reasoning model built by post-training Nvidia's Nemotron 3 Super on synthetic data drawn from roughly three decades of business knowledge and tuned for long-running agents that execute multiple tasks. Salesforce says Koa matches or exceeds leading model performance on CRM actions with three times fewer errors on its own CRM benchmark. Three more pieces shipped with it: AIforce, which answers questions, runs queries, and builds dashboards by reading your Salesforce instance directly; Claudeforce, the Salesforce in Claude beta with 37 pre-built sales skills for deal review, pipeline hygiene, prospect research, and account management, where one admin connects it once and permissions are inherited from existing Salesforce roles; and Headless 360, which exposes the platform through APIs, MCPs, plugins, and skills so customers can work outside salesforce.com. Skills for service, marketing, and commerce teams are planned for later this year. Marc Benioff's line at the keynote: the SaaS-pocalypse was not about the end of software, but it may be about the end of software that makes humans do all the work. ^{[1] [2] [3] [4]}
• Labs agree on third-party testing, and the enforcement question is still open. Anthropic, Google, and OpenAI have discussed creating an industry standards body that tests advanced models before deployment, a conversation CNN reports started with Demis Hassabis' July essay pitching a US-led regulator modeled on FINRA. OpenAI is backing the FRONTIER act, a bipartisan House proposal that would require frontier labs to embed outside evaluators, and embedded evaluators are part of Dario Amodei's plan for giving third-party teams employee-like access to verify safety commitments. Elon Musk called for labs to test each other's models, naming OpenAI, Anthropic, Google, Meta, and three or four leading Chinese companies; Zuckerberg argued labs should slow themselves when safety requires it and that trust and alignment will read as product advantages rather than compliance work. Miranda Bogen, director of the governance lab at the Center for Democracy and Technology, told The Deep View that outside testers can spot problems but cannot change company behavior on their own. She lists what is missing: measurement standards, a channel for reporting failed safety checks to outside stakeholders, mandatory fixes, and a clear allocation of responsibility so evaluations do not become a checkbox exercise. The Deep View's own read is that the Trump administration has said it does not see the risks the industry describes, and getting Chinese labs to cooperate adds a second obstacle. ^{[5] [6] [7] [8] [9] [10]}
• Odyssey-3 runs robots, cars, and drones off one frozen backbone. Odyssey released Odyssey-3, a single model that controls robots, humanoids, cars, drones, and video game characters. The architecture keeps one shared pretrained backbone frozen and trains small task-specific adapters on top, so a new robot or vehicle costs a lightweight layer instead of a full training run. Reported capabilities: control across those five platforms from one model, interactive simulated worlds where agents learn by doing, 77% of real-world driving performance on Indian roads for policies trained purely in simulation, and training loops where AI-generated worlds train the next model. Public access is expected through the developer portal within weeks, so the simulation-to-real numbers are the claim to check against your own hardware. ^{[11] [12]}
• ChatGPT's co-creator shipped a model for decisions instead of prose. Diogo Almeida, who worked on the research behind ChatGPT, spent two years building what TypeSafe calls System One models, and the first public one is JEV. It takes structured questions and returns typed answers plus calibrated probabilities rather than generated text. TypeSafe reports 70 to 500 millisecond answers, 20 to 200 times faster and 40 to 400 times cheaper than comparable LLM workflows, with training built on Reinforcement Learning for Calibrated Decisions so the confidence score means something to calling software. The cost figure TypeSafe is promoting: roughly $42 per billion tokens of equivalent workload. The pitch is the moment where a workflow needs a decision rather than an explanation, like flagging fraud, escalating an alert, or triaging 1,000 records. TypeSafe is taking access requests, so the speed and cost claims are still untested in production. If they hold, a common agent pattern changes: you stop asking a chat model for JSON and writing code to check whether the JSON is sane. ^[13]
• Gemini 3.8 Live adds real-time vision and 97-language voice. Google DeepMind shipped Gemini 3.8 Live, its newest live dialogue model, with real-time vision input, voice support across 97 languages, and upgraded parallel reasoning. Live Extended Thinking adds a heavier mode that reasons for longer before it speaks. Both newsletters flagged it as the same day's release, and the practical difference from the previous generation is that vision and voice arrive in the same live session instead of separate tools. ^[14]
• OpenAI talked pre-IPO money at $1.2T while the Gates Foundation put $1B into access. OpenAI is in early talks for another funding round at a valuation of roughly $1.2 trillion, after annualized revenue topped $40 billion, per the Financial Times. The Gates Foundation pledged $1 billion to expand global AI access. Two smaller rounds landed the same day: Profound raised $180 million at a $1.8 billion valuation after 3x revenue growth in six months and more than 1,000 enterprise customers for AI-search optimization, and manufacturing AI firm CADDi raised $114 million at $1.2 billion. ^{[15] [16] [17] [18]}
• Two teams put models next to the physical world: a 284-pound humanoid and a model wired to lab benches. Agility Robotics unveiled Digit 5, a 5'11", 284-pound humanoid built to work beside people without safety barriers, carrying up to 50 pounds and recharging in nine minutes. Periodic Labs connected its 1T-parameter Neon model directly to physical materials experiments, so the model proposes work, the lab runs it, and the results train the next round. Digit 5 removes the cage between robot and worker; Neon removes the human handoff between hypothesis and experiment. ^{[19] [20]}

Deeper Cuts

• Models pick up personalities from the fiction in their training data. A new study found that LLMs absorb personality traits from fictional human characters in their training data. The mechanism matters more than the headline: if a character's voice appears often enough in pretraining, some of that voice shows up in outputs that were never meant to sound like anyone in particular. The paper is the place to check which traits were measured and how the effect was isolated. ^[21]
• RL helps most on easy problems, and a mini-paper unifies two ways of framing goals. One paper argues that reinforcement learning training mostly improves LLMs on easy problems and proposes a method to fix the gap. Another mini-paper unifies goal-based and hierarchical reinforcement learning under hidden Markov models, which is the theory version of the same complaint: if the goal and the subgoal structure are modeled separately, the training signal gets weaker than it needs to be. Both are research claims to read before you rebuild a pipeline around them. ^{[22] [23]}
• Sakana AI trained 1,000-layer networks without backpropagation. Sakana AI's PC-ALM trains 1,000-layer neural networks using local-only dynamics, which means each layer updates from information it can see nearby instead of a global error signal sent backward through the stack. Backpropagation's memory cost is what keeps networks from going deeper, so a working local-learning method is the interesting part of the claim. The repo is public. ^[24]
• 1,393 subagents refactored a million-line codebase for about $25K, and humans still caught regressions. Nous Research ran 1,393 Fable subagents across a million-line codebase, finishing the refactor in roughly 19 active hours for about $25,000. Human review still caught regressions, which is the number worth keeping next to the price. It puts a real cost on the pattern of handing a large mechanical job to a swarm and reviewing the output instead of writing it. ^[25]
• 74% of surveyed scientists saved time with AI, and verification became the bottleneck. Google's analysis of AI in science found 74% of surveyed scientists saved time with AI, averaging about 6.9 hours a week. The same survey surfaced verification and untested hypotheses as the new constraints: when generation is cheap, the work moves to checking whether the generated thing is true, and that checking has no matching speedup. ^[26]
• Slop grenades: the cost of cheap generation lands on the reviewer. Shopify's Tobi Lütke named the pattern of dumping AI output on coworkers and making them spend the time you saved reviewing it. His argument is that as generation gets cheaper, judgment gets more valuable, and the person who ships unreviewed volume is exporting work rather than doing it. Useful framing for anyone setting review norms on a team this quarter. ^[27]
• Gradients: an essay on tools with a working exhibit. A short piece on gradients that includes an interactive exhibit rather than only screenshots. Worth a pass if you have argued about how much of a design tool's default behavior a designer should accept, or about where the line sits between a tool teaching a technique and a tool hiding one. ^[28]
• Rauno's fifth field note: delete the arrows, stack the cards. Rauno Freiberg's write-up on removing pagination arrows in favor of interactive card stacks, and how that changes what fits in a tight space. The move worth stealing: when a layout runs out of room, the fix is often a different interaction rather than a smaller version of the same one. ^[29]

Tools and Launches

• LongCat-Video-Avatar 1.5. Meituan's MIT-licensed model turns a portrait photo plus an audio file into a lip-synced talking video, with two characters in one scene driven by separate audio tracks, 480p or 720p output, stable identity across long clips, and a text-only mode that generates a character with no photo. The v1.5 fix is lip sync: it swaps in Whisper-Large for the audio processor, so mouth movement stops drifting after a few seconds. Self-hosting needs a 40GB GPU and runs about 44 seconds of compute per second of video. Weights are on Hugging Face. ^[30]
• Meta One. Meta's subscription bundle across Instagram, Facebook, WhatsApp, and Meta AI, with higher AI usage limits and more than 50 features. Meta says its plans have reached 15 million subscriptions and trials. ^[31]
• Gensyn open-1b. A model shipped with a verifiable training record, so outsiders can rerun parts of the training on different hardware and check that it was trained the way Gensyn says. An audit trail for training runs is the piece most model-provenance claims are missing. ^[32]
• Aside. An agent that works across the websites you are already logged into and your local Windows files, with credentials scoped to the task and a confirmation step before high-risk actions like posting or payments. ^[33]
• Sketchpad Live. Turns GPT-Live or Astra into a whiteboard teacher that draws, moves shapes, and narrates a lesson while you follow along. Useful for testing how a spoken model behaves when it also has to keep a visual state consistent. ^[34]
• Concat. An open-source desktop video editor in the CapCut shape: multi-track editing, offline Whisper captions, local text-to-speech, and no account or watermark. The local captions are the part that matters if you edit interview footage you would rather not upload. ^[35]
• OpenArtifacts. A publishing surface for Codex, Claude Code, Hermes, Pi, and OpenCode that turns agent output into HTML or Markdown artifacts you review in a browser, which beats scrolling a terminal transcript when the thing you need to judge is a layout. ^[36]
• Opusfived. A short interactive comedy about agentic coding assistants that cannot do the thing you asked. The premise is one instruction: make one button blue, change nothing else. Worth three minutes if you have ever received a 400-line diff in response to a one-line request. ^[37]
• Sidebranch. Reviews pull requests from inside your running app. You switch branches from an in-page widget and compare worktrees side by side, so visual changes get judged in the real interface instead of from a screenshot in a PR thread. ^[38]

Prompt of the day