ChatGPT connects to bank accounts & Assistants get more action controls - AI News (May 19, 2026)

OpenAI has launched a preview of a personal finance experience inside ChatGPT for Pro users in the U.S. The idea is simple: connect bank, credit card, loan, and investment accounts via Plaid, get a consolidated dashboard, and then ask questions grounded in your actual spending and balances. OpenAI is stressing guardrails—no full account numbers, no ability to move money, and controls to disconnect and delete synced data—because once an assistant can see your financial life, the stakes jump fast.

On the broader “assistants that do things” front, Google is rolling out a new “Thinking level” control in the Gemini app, letting some users choose between faster answers and deeper reasoning. At the same time, Google’s documentation points to more third‑party connectors—think services like design, shopping, and reservations—pushing Gemini toward being an action hub, not just a chat box. And on the OpenAI side, there’s reporting that Codex computer control could expand so tasks can run even when a Mac is locked or asleep—useful, but potentially a major security and platform-policy friction point.

OpenAI is also reportedly tightening the perimeter around voice cloning. The New York Times says OpenAI acquired Weights.gg, a small startup whose public catalog included voice models imitating celebrities and public figures. The hosted service shut down before the deal surfaced, and the team reportedly dispersed inside OpenAI—suggesting this was as much about removing a risky voice library from circulation as it was about building a new product. The big theme here is consent and rights management: the tech is widespread, but the legal and reputational blast radius comes from recognizable voices and misuse at scale.

Zooming out, analyst Benedict Evans argues generative AI is the next platform shift on the scale of PCs, the web, and smartphones—driving a massive reallocation of capital into compute. But he also points out the industry is nowhere near equilibrium: GPU supply, power availability, and data-center buildouts are hard constraints, and model pricing power may weaken as frontier models converge. In a similar vein, Sriram Krishnan says agentic workloads are pushing infrastructure into awkward territory—long-running, context-heavy sessions that stress GPUs, CPUs, memory bandwidth, and especially HBM—setting up opportunities for inference-optimized hardware and new architectures.

On the model side, long context is becoming the battleground, and efficiency tricks are piling up. Nous Research introduced “Lighthouse Attention,” aiming to cut the quadratic attention cost that makes long-context pretraining punishingly expensive. Their pitch is pragmatic: select a smaller set of relevant tokens, then run standard optimized attention kernels on that gathered subsequence—so you get speedups without rewriting the whole stack. Meanwhile, a broader review of open-weight releases highlights the same direction: KV-cache reductions, selective attention budgeting, and compressed attention schemes—plus a revival of local experimentation like activation “steering,” as new forks make it easier to tinker outside hosted APIs.

If all that sounds hard, it is—and a separate set of notes making the rounds argues large-scale pretraining remains surprisingly fragile. The claim is that subtle choices can quietly break real-world behavior: routing or token-dropping tricks that “break causality,” numerical precision errors that bias gradients, and mismatches between training and inference engines that introduce off-policy weirdness. The takeaway isn’t a tidy checklist. It’s that as models scale, the failure modes multiply, and systems engineering details become model quality details.

Meanwhile, teams trying to run AI agents in production are discovering the unglamorous part: reliability and cost. A preview of Inngest’s 2026 benchmark report describes a “confidence paradox,” where only a small slice of teams feel ready to scale their AI systems a few times over, and observability keeps coming up as the biggest gap. In parallel, developers are doing hard math on token economics—like a prompt-caching analysis that finds a surprisingly concrete break-even window for keeping a cache alive. And open-source tooling like Headroom is betting that compressing what you send to the LLM—without permanently losing the ability to fetch originals—can improve both throughput and stability when context windows become a bottleneck.

In coding, Anthropic says Claude Code is now being used across enormous monorepos and legacy systems, and the lesson isn’t “just buy a model.” It’s that you need a harness: curated context files, automated checks, and integrations like LSP and internal tool connectors so the agent can navigate like an engineer without relying on stale indexes. At the same time, open source is feeling the downside of AI scale: the Archestra maintainers say they were swamped by low-quality AI-generated issues and PRs, and they’ve moved to a strict whitelisting workflow just to keep signal over noise. It’s a reminder that “more contributions” isn’t the same as progress.

Finally, two stories underline how contested AI adoption is getting. University of Washington researchers proposed a study using body-worn cameras on preschool teachers to collect first-person classroom footage for AI training—reportedly framed as opt-out for parents. That lands right in the center of consent and surveillance concerns, especially with minors. And the Wall Street Journal reports a broader U.S. backlash: public figures getting booed over AI, anxiety tied to jobs and energy costs, and anger spilling into politics. Add in the ongoing debate about AI-driven wealth concentration in places like San Francisco, and it’s clear the legitimacy question—who benefits, who pays, and who decides—is becoming as important as the next model release.