
Rumors of diminishing returns from bigger models miss what is actually happening: the axes of scaling are shifting from parameters toward data quality, test-time compute, and multi-agent orchestration.

After two years of scarcity, top-tier GPUs are shipping on reasonable timelines again. The next constraint is measured in megawatts, and it is much harder to fix.

A new category of browser is emerging where the address bar is a task box. Early versions are rough, but the interaction model may be more important than any single product.

Retrieval-augmented generation is presented as a hallucination fix. It is not, by default. Here is what actually works — and what does not — after four years of production experience.
Frontier models, benchmarks, and the science behind them.
Deals, funding, and the business of artificial intelligence.
Launches, reviews, and how AI ships to real users.
Building with LLMs, agents, and modern ML infrastructure.
Regulation, safety, labor, and the human stakes of AI.

After a decade in the shadow of pure deep learning, symbolic reasoning is being welded back onto neural systems — not as a rival paradigm, but as scaffolding for reliability.

Learned simulators of physical and social environments are becoming the substrate that reasoning, robotics, and planning increasingly depend on — and the labs that own the best world models may own the next decade.

Rural counties that had never heard of hyperscale compute two years ago are approving multi-billion-dollar campuses. The land, water, and power arithmetic behind the rush.

Users are talking to their AI assistants more and typing to them less. The product implications — for form factor, latency budgets, and interface design — are only starting to sink in.

For years diffusion belonged to images, video, and audio. A new wave of research suggests the same denoising machinery may soon reshape how large language models generate text.

The vector database category spent its adolescence trying to be a standalone product. In 2026 it is quietly becoming a feature of every serious database — and that is a healthier outcome for the people actually building things.

After a decade in the shadow of pure deep learning, symbolic reasoning is being welded back onto neural systems — not as a rival paradigm, but as scaffolding for reliability.

Learned simulators of physical and social environments are becoming the substrate that reasoning, robotics, and planning increasingly depend on — and the labs that own the best world models may own the next decade.

For years diffusion belonged to images, video, and audio. A new wave of research suggests the same denoising machinery may soon reshape how large language models generate text.

Rural counties that had never heard of hyperscale compute two years ago are approving multi-billion-dollar campuses. The land, water, and power arithmetic behind the rush.

Total compensation for senior AI researchers has decoupled from the rest of tech. What is actually being priced — and who is paying.

After two years of scarcity, top-tier GPUs are shipping on reasonable timelines again. The next constraint is measured in megawatts, and it is much harder to fix.

Users are talking to their AI assistants more and typing to them less. The product implications — for form factor, latency budgets, and interface design — are only starting to sink in.

After twenty years of underwhelming rollouts, internal search inside large organizations is quietly becoming useful — driven less by better ranking and more by a shift in what search is expected to do.

The first generation of standalone AI wearables underperformed expectations. The next wave is quieter, more focused, and — perhaps — actually useful.

The vector database category spent its adolescence trying to be a standalone product. In 2026 it is quietly becoming a feature of every serious database — and that is a healthier outcome for the people actually building things.

Teams keep asking whether to fine-tune or retrieve. The honest answer is that the question is malformed — most production systems need both, and the interesting choice is which parts to route where.

Getting an LLM to return valid JSON is the easy part. Making it return valid, safe, on-schema output at scale is a systems problem that most teams underestimate.

The bloc's landmark law has moved from theoretical to operational. The lessons for teams building outside Europe are more direct than most realize.

After years of voluntary commitments and incompatible schemes, a workable set of content-provenance standards is coming together. The question is whether adoption catches up before the political urgency passes.

The predicted deepfake apocalypse did not arrive. The subtler AI influence on elections did — and the lessons for the next cycle are not the ones commentators expected.
Our running dossier on the labs, models, and mechanics defining the frontier this year. Explainers, deep reads, and the numbers that actually matter.

Rumors of diminishing returns from bigger models miss what is actually happening: the axes of scaling are shifting from parameters toward data quality, test-time compute, and multi-agent orchestration.

Every year a new suite of hard reasoning benchmarks lands, and every year a model clears them within months. What that pattern is really telling us about progress.

Mixture-of-experts models were an academic curiosity a decade ago. They now underpin most frontier systems — for reasons that have as much to do with GPUs as with intelligence.

After two years of scarcity, top-tier GPUs are shipping on reasonable timelines again. The next constraint is measured in megawatts, and it is much harder to fix.
“The bitter lesson has not been repealed. It has been generalized. Compute still wins — we are just learning where to spend it.”
Every Monday, Signal&Stack distills the most important developments in AI research, industry, and policy into a single briefing. No hype, no spam — just the signal.
Free. Weekly. Unsubscribe anytime.
Signal&Stack is an editorial publication covering AI — the research, the industry, the products, and the policy questions that follow. We write for professionals who need to understand what matters, why it matters, and what comes next.