Ask ten researchers what AGI will look like and you get eleven answers, most of them hedged. Ask them what changed in the last eighteen months and the answers converge fast. The interesting shift isn’t a single breakthrough. It’s that the field has quietly reordered its priorities: less brute force on pretraining, more effort on thinking at inference time, and a scramble to make models actually do things rather than just describe them.
Here are the AGI technology trends that are doing real work right now, and the ones that still look like a demo waiting for a business case.
Test-time compute replaced raw scale as the headline
For years the recipe was simple: more parameters, more tokens, better loss. That recipe hasn’t stopped working, but the returns per dollar have flattened enough that labs went looking elsewhere. They found it in inference.
Reasoning models spend extra compute before answering, generating long internal chains of thought, backtracking, trying approaches and discarding them. OpenAI’s o-series made this mainstream; DeepSeek’s R1 proved it could be done in the open and reproduced for a fraction of the cost. The practical consequence is that you now buy intelligence with two dials, not one. A model can be small and cheap for routine work and expensive and deliberate for the hard 5% of queries.
That has knock-on effects across the stack. Serving infrastructure has to handle wildly variable latency. Evaluation gets harder, because the same model gives different quality depending on how long you let it think. And pricing pages get messy when a single request can cost a hundred times more than its neighbour.
Why this matters for AGI timelines
If capability scales smoothly with thinking time, then a lab with a fixed model can keep squeezing gains without a new training run. It also means the gap between the best model and the tenth-best model is smaller than benchmark tables suggest, since much of the visible difference is just spend.
Pretraining economics have gone strange
DeepSeek-V3 was trained for a reported $5.6 million in compute, a number that would have sounded like a typo in 2023. Even accounting for the caveats around total R&D spend, the direction is unmistakable: efficient architectures, aggressive mixture-of-experts routing and better data curation are compressing the cost of reaching a given capability level. If you want the fuller picture of where that leaves the frontier, this breakdown of the current state of AI in charts is a good place to start.
The flip side is a frontier that costs billions. The top labs are still spending enormous sums on clusters, and the models at the absolute edge are not cheap to reach. So the distribution is splitting: a crowded, fast-moving middle where capability is cheap and commoditised, and a narrow top where it is extraordinarily expensive.
Open weights keep eating the previous generation
The pattern has become almost mechanical. A closed lab ships a frontier model. Six to twelve months later an open-weight model matches it. Llama, Qwen, Mistral and now the DeepSeek and Kimi lines have all followed this curve, and the lag is shortening, not lengthening. The emergence of competitive Western open-weight efforts built on the same playbook, covered in this look at the so-called DeepSeek of the West, suggests the open ecosystem is no longer a lagging indicator but a parallel track.
For anyone building products, this changes the calculus entirely. Fine-tune an open model on your own data, keep it on your own hardware, and you own the weights, the latency and the bill. The trade-off is operational competence: open weights are only cheap if you have the people who can run them.
Agents are the bet everyone has placed and nobody has cashed
Every major lab now ships some form of agent product. Reliability is the sticking point. A system that completes 85% of tasks sounds impressive until you remember that a five-step workflow at 85% per step succeeds about 44% of the time. Compounding failure is brutal.
What’s actually working today looks less glamorous than the demos:
- Narrow, tool-rich agents that call a handful of well-documented APIs and nothing else, with hard guardrails on each call.
- Coding agents, because the environment gives clean feedback. Tests pass or they don’t, and that signal is worth more than any reward model.
- Human-in-the-loop pipelines where the agent drafts and a person approves, which quietly delivers most of the value without the catastrophic failure modes.
What’s still shaky: long-horizon tasks, anything requiring careful state tracking across hours, and agents that need to recover gracefully from their own mistakes. The 2025 trend is less about smarter agents and more about better scaffolding around mediocre ones.
The interface layer is becoming the real battleground
Model quality is converging. Distribution isn’t. That’s why so much energy is going into making models reachable from wherever work already happens: inside an editor, a terminal, a CRM, a spreadsheet. Standardised tool protocols have done a lot of the plumbing work here, and the effect is that models are turning into infrastructure you call rather than destinations you visit. Anthropic’s push in this direction, summarised in every machine is about to speak Claude, is a decent illustration of where the layer is heading.
The strategic implication is uncomfortable for anyone whose product is a chat box with a subscription. If the model is a commodity and the interface is the moat, then the winners are whoever owns the workflow, the context and the switching costs.
Public patience has become a technical constraint
Something shifted in the last two years that doesn’t show up in benchmark tables: the social licence is thinning. Data centre protests, lawsuits over training data, angry picket lines outside executives’ homes and a steady drumbeat of polling showing public anxiety rising faster than public enthusiasm. The anti-AI backlash reaching Sam Altman’s doorstep is the visible tip of something broader.
This is now an engineering input, not just a PR problem. It shapes where you can build, how fast you can permit a facility, what data you can use, and how much friction regulators apply. Labs that treat safety and transparency as a compliance afterthought are storing up cost for later.
The numbers worth watching over the next year
Ignore the AGI timeline predictions. They’re mostly marketing with a confidence interval attached. Track these instead:
- Cost per million output tokens at fixed capability. This has fallen by more than 99% in under two years and is the single best proxy for how fast the technology is diffusing.
- Benchmark saturation rate. MMLU is dead, GPQA is wobbling, SWE-bench is heading the same way. How quickly a benchmark dies tells you more than the scores on it.
- Task horizon length, meaning how long a task an agent can complete without human intervention. This is the metric that actually correlates with economic impact.
- Energy per useful token. Efficiency gains here determine whether the compute buildout is sustainable or a bubble inflating on borrowed power.
None of these will make a good headline. All of them will tell you more about where AGI research is genuinely heading than another demo video of a robot folding laundry. The field’s real trajectory is being set in inference clusters and efficiency papers, not on stage.