AI in the News

AI in the news: week of October 5, 2025

California signs SB 53, the first US frontier-AI law. Anthropic ships Sonnet 4.5 with an Agent SDK. OpenAI ships Sora 2 with a biometric-scan social app. The AI-layoff narrative consolidates. My take on a heavy news week.

Sid Smith

05 Oct 2025 • 7 min read

What this week actually changed: the first US frontier-AI law got signed, two labs pulled the agent-platform story another step away from "you can run this anywhere," and the AI-cited layoff narrative kept building. This is the first installment of what'll be a regular Sunday post, what happened in AI, with sources, plus what I make of it. The four positions I keep returning to in the longer essays, that AI shouldn't be used as cover for premature workforce cuts, that sensitive data doesn't belong in public AI, that governance is the work, that distributed beats concentrated, show up here as the lens, not the main event.

The governance story: SB 53 lands

Here's the part I want to start with, because it's the biggest piece of governance news in a year. On Monday September 29, Governor Newsom signed SB 53, the Transparency in Frontier AI Act. The first US law that points at frontier AI specifically. Labs training models above 10^26 FLOPs have to publish detailed safety frameworks, report serious incidents inside 15 days, and protect whistleblowers who flag catastrophic risk. Penalties cap at $1M per violation. Takes effect January 1, 2026.

This is the bill that came back after SB 1047 was vetoed last year. Most of the controversial liability machinery is stripped out. The scope is narrowed to the largest model trainers. The focus is transparency rather than gating. It's a meaningfully smaller bill than SB 1047 was, and I think this is the right shape. Governance for frontier AI cannot wait for federal action that isn't coming. State-level transparency-and-reporting requirements are how the framework forms while the federal conversation stalls. The 10^26 FLOPs threshold actually hits the labs whose decisions matter most without sweeping in startups or fine-tunes. The $1M cap is small enough not to be the story and real enough to require actual compliance machinery.

The criticism worth taking seriously is "transparency-only laws don't change behavior, labs will publish frameworks and report incidents and the actual safety work won't move." I don't fully buy it. Putting safety frameworks in the public record changes what labs commit to internally because they know it's going to be read, and the audit trail matters even when the audit doesn't have teeth yet. SB 53 is a foundation, not a finish line, and foundations are how this works. What I'm watching: how the labs write their first frameworks (due after Jan 1), whether other states copy the model, and whether the EU AI Act and SB 53 converge or diverge on reporting. The convergence question has the biggest practical consequences.

The agent-platform front: Sonnet 4.5 plus an SDK

The same Monday, Anthropic shipped Claude Sonnet 4.5, coding and agent-focused, pricing held at $3/$15 per million tokens. Same release brought Claude Code 2.0 features and the Claude Agent SDK as separate launches. The day-one developer reviews were generally positive, faster on long-context coding, real refinement over Sonnet 4.

Two things worth noting on this release. First, the cadence. Sonnet 4 to 4.5 on a half-step number rather than a major version means Anthropic is treating this as refinement, not a generation jump. The capability deltas are real but bounded. Second, the Agent SDK alongside the model is the more strategically interesting move. The play is "Claude as the foundation for your agents" rather than "Claude as the chat product," and the SDK is the day-to-day version of that.

Worth being honest: I run Sonnet 4.5 in Claude Code most days. It's good. The coding workflow improvement is real. The push-back I keep landing on isn't the model or the SDK, it's the assumption that "your agent runs on hosted Claude" is the only architecture worth considering. The Agent SDK works fine against local LLM endpoints once someone wires it up. The mental model that hosted-frontier-model is the default and local-model is the exception is a positioning, not a technical fact.

Sora 2 and the biometric-scan trade

Tuesday September 30, OpenAI shipped Sora 2, synchronized audio, sharper physics, multi-shot consistency that's the first time I've watched AI video without obvious tells. Released alongside it: a TikTok-style iOS app called Sora, invite-only in US/Canada. The model itself is impressive. The audio sync, the physics fidelity, the consistency across cuts. Sora 2 is the first AI video model where I can imagine reasonable consumer-grade output without obvious tells.

The feature I want to be loud about is "characters." Users record a short video and audio clip of themselves to enable dropping their likeness into Sora-generated scenes. This is, structurally, the most aggressive sensitive-data-into-public-AI move I've seen from a frontier lab. The pitch is fun and creative. The underlying trade is "give us a high-fidelity scan of your face and voice in exchange for being in the AI videos."

I'm not going to use it. I'd recommend most people don't either. The utility is being in fun AI videos. The cost is OpenAI now has biometric data on you sufficient to generate arbitrary video. That's a trade I would not make, and the framing actively hides the cost side of the transaction. Biometric data sent to a hosted AI service is not getting unsent, and the company holding it has commercial incentives that don't perfectly align with the person whose face it is.

The TikTok-style social feed is the second concern. The pattern of "user-generated AI content in an algorithmic feed" reproduces every problem the existing short-form video platforms have, plus the new ones AI introduces. Deepfake misuse, consent on third-party likenesses, scaled disinformation, all the same problems with a more capable engine. I don't think this product is going to age well in its first incarnation. Worth watching the consent and impersonation incidents over the next 60 days.

DevDay anticipation

OpenAI DevDay is Monday October 6 and the pre-event coverage is heavy. Expectations going in, per InfoQ's preview and the community boards: GPT-5 Pro API availability, AgentKit (visual agent builder), Apps SDK (third-party apps inside ChatGPT), Sora 2 in the API. I'll cover the actuals in next week's roundup. The frame I'm watching going in is the platform-vs-product split. OpenAI is increasingly pitching itself as the platform layer for agents. AgentKit, the Apps SDK, the 800 million weekly users that make ChatGPT itself the distribution channel. That's a different bet than "we sell tokens" and a different competitive surface than the model-quality conversation. The platform play makes lock-in much stickier and the data flow much more centralized.

The slow-burn story: AI-cited layoffs

Salesforce's Marc Benioff said in early September that the company has cut customer service from 9,000 to 5,000 with AI agents covering the gap. The narrative kept building through September. Klarna, IBM, smaller firms, the public framing of "AI replaced the workers" is becoming load-bearing for executives explaining why the headcount went down.

I want to be plain about where I sit on this. The displacement is real and it's accelerating faster than I expected. I always knew it was coming. I've spent my entire career on systems automation, and the through-line of that career is that work which can be automated eventually gets automated. The Salesforce story is a real story. What I keep coming back to isn't whether AI can do the work; it's the pace.

Short-term incentives are driving the pace. Companies aren't cutting because the AI is ready. They're cutting because the AI narrative is convenient and the markets reward the cuts. That's the part I push back on, not the displacement itself, but the rush to do it before the human-and-AI workflows have been figured out. Cuts in 2025 to chase the next earnings call are different from cuts in 2027 after the team has actually rebuilt the work around the tools.

The sustainable shape is human+AI collaboration, and the firms that figure out the collaboration will outperform the ones that just cut. To be clear: the headcount still shrinks under collaboration. It just shrinks less and shrinks well. The work product is better, the operational gaps are smaller, and the 2027 correction cost is lower. Companies optimizing for headline cuts in this earnings cycle are buying the opposite trade. I'd rather be wrong about the pace than be caught off guard. Hope it's slower or smaller than the realistic view says. Plan for it being faster.

Quieter items

A few smaller pieces worth a line. The Dragon Hatchling (BDH), a biologically-inspired LLM architecture from a Polish research team, hit arXiv on September 30. Scale-free network of locally-interacting neuron particles. Probably nothing in the short term; long term, the architecture-diversity question matters, the field over-converged on transformer-shaped models and the alternatives are getting more interesting again. Anthropic's Imagine-with-Claude image-generation preview extended another week and opened to Pro users; still gated, still preview-only. And Google's September roundup shipped. Gemini 2.5 updates, Veo 3 video improvements, Workspace integrations. Quiet month from Google relative to OpenAI and Anthropic. The cadence will swing back.

What this week tells me

Three takeaways. The governance ceiling moved. SB 53 is the first US frontier-AI law, and whatever you think of its specifics, the precedent matters more than the content. The next state law will start where this one ended. The federal version, if and when it comes, will use SB 53 as a reference.

The hosted-frontier-AI bet doubled down. Anthropic's Agent SDK, OpenAI's Sora app and the anticipated Apps SDK are both labs building the platform layer that locks the agent stack to the hosted model. The principled response is to build the same agent stack against local models so the lock-in is optional rather than structural.

And the labor narrative is consolidating around "AI replaces workers" faster than I expected. The displacement is real. The pace is what's wrong, and the financial logic is the driver. The firms figuring out human+AI collaboration will outperform the firms racing to cut headcount on the next earnings call, but the headcount shrinks either way. Next Sunday: DevDay actuals, the Sora app's first ten days of incidents, and whatever else lands midweek.