Two weeks away from your feed and the AGI research news cycle has already lapped you twice. A new reasoning model tops a leaderboard, a lab drops a safety framework, a paper claims an emergent capability that turns out to be a prompting artifact, and somewhere a founder says “we’re basically there” for the fourth time this year. Following it all is genuinely hard work, and most coverage doesn’t help, because it treats every announcement as equally seismic.
The useful version is narrower. Here’s a working map of where AGI research actually stands, which results have survived scrutiny, and where the field is quietly stuck.
What Actually Qualifies as AGI Research News
Plenty of things tagged “AGI” are not research. A funding round is not research. A rebranded product tier is not research. A ninety-second demo video with a tight crop and no failure cases is not research.
Real research news tends to look dull by comparison. An ablation study showing which component carried the gain. A scaling curve that bends where the authors predicted it would. A benchmark with a genuinely held-out test set that nobody trained against. When a lab publishes something like that, it usually gets a fraction of the attention of a product launch, and it usually matters more six months later.
Three Fronts Where the Progress Is Real
Reasoning, and the shift to test-time compute
The biggest structural change of the past two years wasn’t a bigger model. It was where the compute gets spent. Instead of only scaling training, labs started spending at inference: generating many candidate solutions, scoring them, searching over them, and letting the model deliberate before answering. OpenAI’s o-series kicked this off publicly in late 2024, and DeepMind and Anthropic followed with thinking modes of their own.
The ARC-AGI leaderboard shows the shift cleanly. In late 2024, OpenAI’s o3 scored 87.5% on the public evaluation set, up from roughly 5% for GPT-4 a couple of years earlier. The catch, widely reported at the time, was cost: about $20 per task in inference compute for the high-compute configuration. That gap between what a model can do when you let it think for twenty minutes and what it does in a chat window is now the central engineering question in the field.
Agents that finish work rather than demo it
On SWE-bench Verified, a benchmark of real GitHub issues pulled from open-source projects, the best scores sat near 2% in early 2023. By 2025 the frontier was clearing 70%. That’s not a marketing number; the tasks are public and independently runnable.
The more instructive statistic comes from METR, the evaluation nonprofit, which measures how long a task a model can complete with 50% reliability. That horizon has been doubling roughly every seven months. Extrapolate carefully here, because the trend is measured on a specific task distribution and reliability is not the same as capability.
Memory and continual learning, still the weak link
Context windows stretched past a million tokens. That’s useful. It is also not learning. A model with a million-token context still starts from scratch tomorrow, and the field knows it. Genuine weight-level adaptation without catastrophic forgetting remains one of the least-solved problems in the discipline, and it gets discussed far less than it should.
Every Benchmark Gets Beaten, Which Tells You Something
FrontierMath launched in late 2024 as a hedge against contamination: unpublished problems, expert-written, deliberately hard. Within months, frontier models were clawing through double-digit percentages. Humanity’s Last Exam arrived in January 2025 with roughly 3,000 questions across a hundred-plus subjects, and models went from single digits to clearing a meaningful slice of it inside a year.
This pattern is worth understanding rather than mocking. As the track record of AGI intelligence tests shows, a benchmark is a fixed target, and fixed targets get hit. Some of the gains come from genuine capability, some from contamination, and some from the fact that researchers get to see the test. Reading a press release that says a model “beat the bar” without knowing which bar, which version, and who ran it is roughly as informative as a horoscope.
Safety Research Moved From Philosophy to Engineering
The most under-covered AGI research news is happening in interpretability and evaluation. Mechanistic interpretability teams have gone from admiring attention patterns to tracing actual circuits inside models and building sparse autoencoders that decompose activations into human-readable features. None of this solves alignment. All of it makes the system less opaque than it was in 2022.
Evaluation infrastructure matured too. Government institutes in the UK and US now run pre-deployment testing on frontier models before release, and third-party audits have become a normal part of launch cycles rather than a concession. Dangerous-capability evaluations, alignment audits, and red-team protocols are starting to look like a real engineering discipline instead of an academic exercise.
Where the Field Keeps Getting Stuck
Strip away the launch cadence and the same four problems keep showing up. Sample efficiency: a child learns a new word from a handful of exposures, while models need orders of magnitude more. Compositional generalization: systems that ace a benchmark frequently stumble when the task is reworded or the setting shifts slightly. Long-horizon reliability: a 70% per-step success rate compounds into failure over a fifty-step task. Grounding: models reason fluently about the world while never having touched it.
These aren’t minor gaps, and they’re the reason researchers who work on them tend to be the most skeptical voices in the room. A closer look at what the field is building makes the pattern clear: capability gains cluster around tasks with verifiable answers and clean reward signals. Everything else moves slower.
How to Read AGI Research News Without Getting Fooled
- Check the benchmark’s birthday. If the test set predates the model’s training cutoff, ask how contamination was controlled.
- Find the baseline. A model scoring 84% means little without knowing what the previous best was and what a simple heuristic gets.
- Look for cost per task. Capability at $20 per query and capability at $0.002 are different products.
- Separate capability from reliability. Succeeding on two out of three attempts is not the same as doing the job.
- Watch what researchers publish, not what executives promise. Negative results and failed replications are the most honest signal in the field.
That filter alone eliminates most of the daily churn. Separating what’s changing from what’s loud gets easier once you know which questions to ask, and the answers are usually sitting in the paper’s limitations section.
What to Watch Over the Next Year
Three things are worth tracking closely. First, the price of reasoning. Test-time compute is expensive now, and every order-of-magnitude drop changes which applications become viable. Second, reliability under long horizons, measured by independent evaluators rather than vendor-reported numbers. Third, whether labs start publishing the work that failed, because a field that only reports wins can’t tell you how close it really is.
Meanwhile, the fundamentals of how these systems are assembled are worth understanding in their own right, since the architecture behind AGI systems explains a lot of the observed behaviour, including the failures. If you want to keep a running list of what’s genuinely worth reading, the honest shortlist of recent AGI developments that hold up is a reasonable place to start.
The through-line is simple enough. Progress is real, concentrated in domains where success can be machine-checked, and slower than the headlines suggest everywhere else. The researchers closest to the work tend to be the least certain about timelines, and that uncertainty is information, not evasion.