GPT helps crack Erdős conjecture & Talent and compute arms race - AI News (Apr 28, 2026)

[0:35] GPT helps crack Erdős conjecture

Starting with that math surprise. A young amateur, Liam Price, posted what looks like a genuine solution to a long-standing Erdős conjecture about “primitive sets” and a particular sum Erdős studied. What’s striking is that the key move reportedly came from GPT-5.4 Pro making an unusual connection—pulling in a known formula from a neighboring area that researchers hadn’t applied in this exact way. Experts like Terence Tao and others say the AI’s raw proof was messy, but the central idea appears to hold up after human reconstruction. The takeaway isn’t “AI replaces mathematicians.” It’s that models can now propose unfamiliar pathways—while humans still carry the burden of rigor, explanation, and trust.

On the research side, Google DeepMind introduced Decoupled DiLoCo, a distributed training approach meant to keep large runs moving even when parts of the system fail or when compute is spread across regions. Instead of tightly locking every accelerator into the same step, it allows looser synchronization, so an outage doesn’t freeze the entire job. The significance is operational: frontier training is increasingly a reliability problem as much as an algorithmic one.

Another paper—nicknamed “Vision Banana”—argues something provocative: image generators can be tuned into strong general visual understanding systems by turning perception tasks into image outputs, like producing a segmentation mask or depth map as an image. If the results hold up broadly, it suggests generative pretraining may become an even more central route to general-purpose vision, reducing the need for separate specialized architectures for every task.

Meta research also surveyed “efficient video intelligence” as of April 2026, emphasizing a practical trend: compressing and distilling video understanding so it works on real devices and long clips, not just short benchmarks. The through-line is efficiency—less redundant processing, smarter temporal handling, and on-device models that are finally credible for tracking and segmentation. It’s a reminder that progress isn’t only bigger models; it’s making them usable where latency, battery, and cost actually matter.

Now to sovereignty—because it’s everywhere in policy decks right now. One critique argues “sovereign labs,” meaning nationally branded frontier-model builders, are mostly unnecessary for typical enterprise needs. The author draws a line between sovereign pre-training and sovereign deployment, and says most companies really want data residency, auditability, and protection against their data being absorbed into someone else’s training loop. That’s less about model nationality and more about controlling data flows and deployments—often using open models locally, with strict isolation for sensitive inputs.

Still, sovereign AI is attracting major alliances. Cohere and Germany’s Aleph Alpha announced a partnership positioned as an independent, enterprise-grade alternative for regulated sectors, with sovereign cloud hosting in the mix. Whether this becomes a real technical advantage or mainly a procurement story will depend on performance, integration, and long-term support—but the demand signal is clear: governments and regulated industries want leverage and options.

A few product and trust stories round out the day. Google appears to be preparing a shift of the Gemini app toward a credit-based usage model. If that lands, it’s a more flexible way to price heavy features—especially long multimodal sessions and agentic tools—while making costs feel more “metered” than “tiered.” Expect this to influence user behavior, because credits change how people experiment.

Canva also dealt with a trust-and-safety mess: users reported its Magic Layers feature was replacing the word “Palestine” in designs. Canva says it fixed the bug and added safeguards. Even if it was unintended, it’s a sharp example of why creators get nervous when AI tools touch existing content: a small, opaque change can become politically loaded instantly, and trust is hard to win back once people fear silent edits.

In governance and public posture, OpenAI published a new “Our Principles” statement, framing commitments around democratization, empowerment, prosperity, resilience, and adaptability—while acknowledging that in some cases it may prioritize safety over maximum user control. These documents don’t settle debates on their own, but they signal how labs are positioning themselves as scrutiny rises from regulators and the public.

Finally, two reality checks—one societal, one operational. A writer argued for “generative AI vegetarianism”: a personal stance of opting out of generative AI tools in daily life to preserve autonomy, craft, and critical thinking, while still allowing older, narrower automation like spam filtering. Whether you agree or not, it’s a useful label for a growing counter-movement against default AI adoption.

And in San Francisco, a boutique called Andon Market is being billed as the first retail store “run by an AI agent,” Luna. The experiment gave Luna money, a lease, and control over decisions—yet early outcomes include bizarre over-ordering, missing price tags, scheduling shutdowns, and a reported operating loss. It’s an unusually honest demo of the gap between persuasive AI interfaces and the messy, physical, exception-filled world. Agents can plan and talk; running a store still demands dependable execution—and humans are quietly doing much of that work.