Ask ten AGI researchers what they’re building and you’ll get a dozen definitions, two of them from the same person on different days. That ambiguity isn’t lazy reporting. It’s the actual state of the field. Any useful AGI research overview has to start by admitting that the target keeps moving, and that a lot of the disagreement is about definitions rather than about machines.
The working version most labs operate with is machine intelligence that matches or beats humans across essentially any cognitive task, including ones nobody thought to train for. That last clause does a lot of work. A system that can beat a grandmaster at chess and still fail to notice it’s being scammed on a phone call isn’t general in any meaningful sense. If you want the baseline mechanics of how these systems learn, this plain-English guide to how artificial intelligence works covers the foundations before the ambition.
“General” Is Doing More Work Than “Artificial”
In 2023 a group at DeepMind published a framework that split the problem into two axes: how well a system performs, and how broad its capabilities are. On the performance axis they laid out levels running from emerging to competent to expert to virtuoso, with artificial superintelligence sitting past the end. The useful move was separating capability from generality. A narrow system can be expert-level at one thing. A general system can be mediocre at everything, which isn’t obviously progress.
Most of the public argument collapses those two axes. Headlines about models “reaching PhD level” usually mean a model scored well on a set of PhD-style questions in physics or chemistry. That’s real, but it says nothing about whether the same model can plan a three-week experiment, order the reagents, notice a contaminated sample, and revise the protocol. The gap between answering questions and doing work is where most AGI research actually lives.
The Four Research Camps, Roughly
Scale-first: let the loss curve do the thinking
This camp bets that general capability is an emergent property of enough parameters, data and compute. The evidence isn’t nothing. Scaling laws held up well enough for years that labs built billion-dollar clusters on the strength of them, and the 2022 Chinchilla result reframed how data and parameters should be balanced (roughly twenty training tokens per parameter, far more data than most teams had been using). The weakness is that scale lowers loss, not necessarily the specific capability you wanted. A model can get better at everything measurable and still not plan.
Architecture and cognition
Another group argues that transformers plus next-token prediction have structural limits, and that you need explicit machinery for memory, reasoning steps, working state, or something like a learned internal world model. DeepMind’s Gato in 2022 showed one network handling over 600 different tasks, but at narrow-specialist competence. More recent work on inference-time search, tool use, and structured reasoning sits in this bucket too. So does the recurring interest in neurosymbolic hybrids, which pair learned pattern recognition with explicit rules.
Embodiment and developmental routes
Some researchers think you cannot get generality from text alone, because concepts like heavy, warm, stuck, or fragile are grounded in physical experience. This camp builds robot learning, simulation environments, and small-scale developmental systems that learn like infants rather than ingesting the internet. Moravec’s paradox is the animating frustration here: skills humans find trivial (picking up a cup, walking across gravel) have proven brutally hard to encode.
Safety-first
A large and growing slice of the field treats alignment as a precondition, not an afterthought. Work on interpretability, scalable oversight, red-teaming, and evaluations of dangerous capabilities now consumes serious headcount inside frontier labs. This camp’s argument is blunt: if you don’t know how to verify what a system is doing internally, building a more capable version of it isn’t obviously wise.
Evaluation Is the Weakest Link in the Whole Enterprise
Measuring general intelligence is genuinely hard, and the field’s benchmarks keep dying. The ARC-AGI suite, built by François Chollet, was designed to resist memorization by testing novel visual puzzles. Frontier models went from around 2% in 2023 to over 80% on the first version by early 2025. The harder second version reset scores back into the single digits and low twenties. That pattern, benchmark built, benchmark beaten, benchmark retired, repeats every eighteen months or so.
Contamination is the quiet problem. When training data is scraped from the web, test questions end up in the training set, and a model that “solves” a benchmark may simply have seen it. METR’s work on task-completion time horizons takes a different angle, measuring how long a task a model can finish reliably. Their estimates put the 50%-reliability horizon at roughly fifty minutes for top agents in early 2025, doubling every seven months or so. Time horizons are more useful than exam scores because they map onto something you can actually delegate.
Agents Are Where the Ambition Meets Reality
The last two years have shifted attention from chatbots to agents: systems that take a goal, break it into steps, use tools, and check their own work. Frameworks for building them have matured quickly, and platforms like SuperAGI and its approach to autonomous agent tooling show how far the developer infrastructure has come. The interesting failures are all reliability failures: an agent that completes 80% of a task twenty times in a row is worth far less than the percentage suggests, because 80% repeated is roughly nothing.
This is also where the gap between research demos and production systems shows. In telecom network operations, for instance, an incident-first blueprint for using AI to handle outages rather than alarm queues is a world away from a general problem-solver, but it’s the kind of constrained, high-stakes deployment where automation either earns trust or gets switched off.
Compute, Power, and the Data Wall
Training compute for frontier models has grown roughly four to five times a year by most tracking estimates, which is a rate nothing in industrial history sustains for long. The constraints now are physical: gigawatt-scale power agreements, transformer supply, water for cooling, and the simple fact that high-quality human text is finite. Synthetic data and reinforcement learning environments are the two main workarounds under active development. Neither has been proven at the scale the field needs.
Money is a constraint of a different kind. A frontier training run now costs somewhere in the tens to low hundreds of millions of dollars, and results are not guaranteed. That economics pushes labs toward incremental wins on products people already pay for, which is partly why general capability research and commercial AI search products like Andi Search, which answers queries directly instead of returning links, are being developed by overlapping teams with overlapping budgets.
Signals That Tell You Something Real
Most AGI commentary is either breathless or dismissive. Filtering for genuine updates is easier if you know what to look for:
- Reliability curves, not peak scores. A system that succeeds 95% of the time on a two-hour task is qualitatively different from one that succeeds 60% of the time, even if both hit the same benchmark high score.
- Transfer to unseen domains. If a model trained heavily on code suddenly gets better at lab protocol design, that’s a generality signal. If it only improves on code, it’s a specialist.
- Sample efficiency. Humans learn new board games from a rulebook. Models still need thousands of examples. Any credible drop in that number matters more than another percentage point on an exam.
- Interpretability results that hold up. Knowing why a model made a decision, verified independently, is a prerequisite for deploying anything powerful enough to be worth the trouble.
- Cost per unit of capability. Declining inference costs change what’s economically viable far more than benchmark charts suggest.
The honest position on AGI timelines is that the field has produced genuine, measurable progress on narrow and increasingly general tasks, alongside a striking inability to define the finish line or verify when it’s been crossed. Those two facts sit together, and neither cancels the other out. What separates serious research from marketing is whether a claim comes with a reliability curve attached, tested on tasks the model couldn’t have memorized, with the failure modes published alongside the wins.