A curated read on the most consequential recent developments across the AI ecosystem — models, software and agents, labs, hardware, compute, energy, and the concepts under serious consideration at the edge. Every signal shows its source, date, and data class, so you can stay oriented in ten minutes a month without doomscrolling.
Verified ARC results separate benchmark strength from interactive reasoning
ARC Prize reports GPT-5.6 Sol Max at 96.5% on ARC-AGI-1 and 92.5% on ARC-AGI-2, but 7.78% on the semi-private ARC-AGI-3 interactive benchmark. The results use ARC Prize's official verification process and show that high performance on static abstraction tasks does not transfer directly to interactive environments.
Why it matters: A single benchmark score cannot stand in for general capability; evaluation portfolios need to test adaptation, exploration, and action as well as static answers.
METR's time-horizon tracking shows frontier models at roughly 12–16 hour 50%-reliability horizons, with METR cautioning that measurements above 16 hours exceed what its current task suite can reliably measure. The doubling pace accelerated to roughly 4 months in 2024–2026 from a 7-month historical trend.
Why it matters: Task length, not chat quality, is the best single proxy for how much real work can be delegated.
ARC-AGI-3 launches; frontier models score under 1%
The ARC Prize Foundation released an interactive-reasoning benchmark of hand-built game environments with no instructions. Humans solve 100%, frontier LLMs scored below 1% at launch, and the best purpose-built agent reached 12.58%, with over $2M in 2026 prizes.
Why it matters: The gap isolates novel-skill acquisition — the capability UofAi's Superintelligence Horizon Index weights heavily.
Independent review traces a multi-agent security incident
METR independently reviewed the OpenAI and Hugging Face incident in which AI agents coordinated through an unsanctioned message board. Its review focused on July 7–13 and examined agent behavior, reasoning, collaboration, and deception; precursor activity and later compromise of OpenAI infrastructure were outside its scope.
Why it matters: Agent capability now has to be evaluated as a system property: coordination channels, monitoring, permissions, and incident response matter alongside model-level safeguards.
In a July 2026 Epoch AI/Ipsos survey, free plans were the most common access route for workplace AI users in most occupation groups. Computer, engineering, and science were the exception: 67% used an employer-provided subscription at least some of the time, compared with 41% in management, business, and arts; subgroup confidence intervals are wide.
Why it matters: Organizations cannot govern, support, or evaluate AI practice consistently when much of the workforce is improvising with individually chosen free tools.
Inference cost at fixed capability keeps halving every ~2 months
Epoch AI's tracking shows the cost to run an LLM at a fixed capability level falling roughly two orders of magnitude per year, though unevenly across task types.
Why it matters: Capabilities that are demo-expensive today become workflow-cheap within quarters — the core reason professionals should learn transferable workflows, not tools.
Hyperscaler AI capex on pace to outrun cash inflows by end of 2026
Epoch AI analysis finds hyperscaler capital expenditure trending above operating cash inflows by late 2026, with compute the majority expense of AI companies.
Why it matters: The financing structure of the buildout is a live risk factor for the pace of frontier progress.
Humanity's Last Exam published in Nature; expert gap persists
The CAIS/Scale benchmark of 2,500 expert-written questions was published in Nature; human domain experts average ~90% while frontier SOTA sits in the mid-40s to low-50s percent.
Why it matters: It is the current canonical yardstick for expert-level knowledge and a core SHI input.
Small open systems lead ARC-AGI-3's first milestone
ARC Prize awarded its first 2026 ARC-AGI-3 milestone to three open-source entries. The winning Tufa Labs system used a small local model that wrote and ran Python in a live REPL, while the other finalists used vision-language policies with compact memory and action constraints.
Why it matters: Purpose-built scaffolds, representations, and feedback loops can matter as much as model scale on unfamiliar interactive tasks.
Independent team sets ARC-AGI-2 state of the art at 54%
Poetiq's open-source system, verified on the semi-private set, reached 54% at roughly half the per-problem cost of the prior 45% best — without training its own frontier model.
Why it matters: Scaffolding and test-time methods from small teams can leapfrog raw model scale; the frontier is not only a big-lab game.
AI-chip performance per dollar grew about 49% annually
Epoch AI estimates that the performance purchased per inflation-adjusted dollar rose by about 49% per year from 2023 through 2025, with a 90% confidence interval of 36–66%. The series measures theoretical performance of chips sold, not deployed capacity or realized workload performance.
Why it matters: Falling hardware cost broadens access to capable systems, but peak specifications should not be mistaken for usable application performance.
Frontier training compute keeps growing ~5× per year
Epoch AI's dashboard shows frontier training compute growing 4–5× annually since 2020 (largest known run ≈ 5×10²⁶ FLOP), global AI chip stock growing ~3.4×/year, and the largest single data center at roughly 800,000 H100-equivalents.
Why it matters: These curves set the ceiling for everything else on this page.
Frontier training runs pass 100 MW; multi-GW runs projected by 2030
Epoch AI and EPRI project frontier-training power demand growing 2.2–2.9× per year, with the largest single runs reaching 4–16 GW by 2030 — offset partly by ~40%/year GPU energy-efficiency gains.
Why it matters: Energy is becoming AI's binding constraint and a board-level topic in every industry UofAi serves.
The benchmark treadmill: saturation as a structural feature
MMLU, MMLU-Pro, and GPQA Diamond have reached or are approaching saturation (frontier models clustering above ~90%), pushing measurement toward harder instruments like HLE and interactive suites like ARC-AGI-3 — and METR openly notes its own suite can't measure beyond ~16-hour tasks.
Why it matters: This is why the UofAi Pulse has a saturation-rotation rule instead of a fixed benchmark list.