TRINITY transferred zero-shot to four unseen tasks (AIME, BigCodeBench, MT-Bench, and GPQA). On average, the evolved coordinator surpassed every individual constituent model in its pool, including GPT-5, Gemini 2.5-Pro, and Claude-4-Sonnet.

In nature, complex problems are rarely solved by a single monolithic entity, but rather by the coordinated efforts of specialized individuals working together.

At 50 million tokens, the design space for AI applications changes fundamentally.

Almost all the evidence points to very fast software progress: each year, the training compute needed to get to the same capability declines several times — possibly even ten times or more.

The depth of recursion becomes a tunable compute axis at inference time, requiring no retraining. A small model, by reading itself, can iterate toward answers that neither it nor any of its workers could reach in a single pass.

The architecture scales horizontally to 300 sub-agents executing across 4,000 coordinated steps simultaneously, a substantial expansion from K2.5's 100 sub-agents and 1,500 steps.

memory-driven experience scaling represents a crucial new frontier for agent scaling

existing TTS methods often discard the exploration trajectory and treat the final answer as the only useful outcome

When a Fugu model is allowed to call itself recursively, reading its own prior output as context and deciding whether to revise its coordination strategy, a new form of test-time scaling emerges.

The depth of recursion becomes a tunable compute axis at inference time, requiring no retraining. A small model, by reading itself, can iterate toward answers that neither it nor any of its workers could reach in a single pass.

Parameters are estimated by unweighted least squares. Time t is measured in years since the first observation in each dataset.

On a 150-class benchmark, the surrogate fully replaces the teacher

The real unlock is compound scaling—token spend grows linearly while output grows exponentially.

At 770M parameters, a looped model achieves the downstream quality of a 1.3B fixed-depth Transformer trained on the same data — roughly half the parameters for the same quality.

They will have to go through our journey above of ingesting data, collecting tribal knowledge, and more – and they will have to do so for each individual customer they work with.

Because small, cheap, fast models are sufficient for much of the detection work, you don't need to judiciously deploy one expensive model and hope it looks in the right places. You can deploy cheap models broadly, scanning everything, and compensate for lower per-token intelligence with sheer coverage and lower cost-per-token.

The capability rankings reshuffled completely across tasks. There is no stable best model across cybersecurity tasks. The capability frontier is jagged.

we can reach the same capabilities with over an order of magnitude less compute than our previous model, Llama 4 Maverick.

scaling Muse Spark with multi-agent thinking enables superior performance with comparable latency.

we can reach the same capabilities with over an order of magnitude less compute than our previous model, Llama 4 Maverick.

We see continued gains from inference scaling on larger projects, suggesting they may be solvable given enough tokens.

On a single H200 GPU with 1.5TB host memory, MegaTrain reliably trains models up to 120B parameters.

we discover a clear scaling law: as the underlying foundation models improve, the quality of the generated papers increases correspondingly.

19世紀の経済学者ジェヴォンズは、蒸気機関の効率向上によって石炭の消費効率が上がると、かえって全体の消費量が増えることを見出しました。

This observation challenges widely accepted test-time scaling laws, leading us to hypothesize that errors within the reasoning path scale concurrently with test time.

Without any architectural modification, MinerU2.5-Pro achieves 95.69 on OmniDocBench v1.6, improving over the same-architecture baseline by 2.71 points and surpassing all existing methods including models with over 200× more parameters.

Sandboxes made for running tens of thousands of agents

But there's one post-American system that's easy to imagine. The project to rip out all the cloud connected, backdoored, untrustworthy black boxes that power our institutions, our medical implants, our vehicles and our tractors; and replace it with collectively maintained, open, free, trustworthy, auditable code. This project is the only one that benefits from economies of scale, rather than being paralyzed by exponential crises of scale. That's because any open, free tool adopted by any public institution – like the Eurostack services – can be audited, localized, pen-tested, debugged and improved by institutions in every other country.

For these individual successes to scale into a continent-wide shift, however, structural barriers must be addressed. The path to digital sovereignty is not a single, grand gesture but a series of deliberate, often difficult, choices. The examples from Austria, France, and the ICC show that the journey begins with a single, courageous step, often prompted by the mundane reality of a data protection assessment.

* "Digital platforms are used for hybrid campaigns."* "EU can't compete with US tech ON THEIR TERMS."* "Post-reality US is what happens when tech is unregulated."* "Ireland is a Trojan Horse for Big Tech."* "The Digital Omnibus is sabotage."

would take seriously the fact that intelligence is now being scaled and distributed through organizations long before it is unified or fully understood

That raises fixed costs and favors scale—paradoxically advantaging the largest platforms that can afford regional bifurcation while crushing subscale competitors.

This will be helped by increased "re-shoring" efforts of many Western Governments to bring "technological sovereignty" to web infrastructure, reducing our reliance on such a small group of American companies. This results in a whole bunch of small, local web hosting services popping up in each country, offering forum hosting with the servers within the country with technical and legal structures especially tailored for their respective markets.

The new litmus test isn’t “Does it scale?” It’s: “Does it spread? Does it take root? Can it compost and regrow?”

support of local and national authorities, local commoners can rely on translocal support structures.

A 2022 paper from Google concludes that making GPT-3-like models bigger makes them more fluent, but no more trustworthy.

f scaling doesn’t get us to safe autonomous driving, tens of billions of dollars of investment in scaling could turn out to be for naught.

scaling starts to falter on some measures, such as toxicity, truthfulness, reasoning, and common sense.

What’s more, the so-called scaling laws aren’t universal laws like gravity but rather mere observations that might not hold forever, much like Moore’s law, a trend in computer chip production that held for decades but arguably began to slow a decade ago.11

There are serious holes in the scaling argument. To begin with, the measures that have scaled have not captured what we desperately need to improve: genuine comprehension. Insiders have long known that one of the biggest problems in AI research is the tests (“benchmarks”) that we use to evaluate AI systems

hat should we do about it? One option, currently trendy, might be just to gather more data.

but the process of identifying trustworthy mailers and notifying them does not scale well to large numbers of small mailers

folgezettel pushes the note maker toward making at least one connection at the time of import.

it is technically impossible to feed the entire population of the planet with organic produce.

Last fall, I spent several days in New York City, during which time I visited a home owned by a group of pacifist Christians that lives from a common purse—meaning the members do not have privately held property but share their property and money. Their simple life and shared finances allow their schedules to be more flexible, making for a thicker immediate community and greater generosity to neighbors, as well as a richer life of prayer and private devotion to God, all supported by a deep commitment to their church.This is, admittedly, an extreme example. But this community was thriving not because it found ways to scale down what it asked of its members but because it found a way to scale up what they provided to one another.

I’ll be speaking with and writing about people working on some of the tools and communities that I think help point ways forward—and with people who’ve built fruitful, immediately useful theories and practices

Abb. 9 Im Normalfall erarbeitete man jedoch eine detaillierte interne Feinsortierung des Belegmaterials häufiger Wörter. Naturgemäß hätte jede Dimension der Analyse (chronologisch, grammatisch, semantisch, graphisch) die Grundlage einer eigenen Sortierordnung bilden können.

Folgezettel

rule-based algorithms like decision trees are not affected by feature scaling.

Isaac and I uh with another colleague we did a little bit of work trying to look at what would the Swedish policy or the UK policy indeed look like if it was carried out globally and it would look at something like two and a half degrees Centigrade of warming if 00:31:58 not more

scaling fractally

It says the world's wealthiest 1% produce double the combined carbon emissions of the poorest 50%, according to the UN.

gsettings set org.gnome.mutter experimental-features "['scale-monitor-framebuffer']"

Scale, inevitably leads to power-law distributed outcomes, leading to the inevitable concentration of talent and resources among a few investigators pursuing a few lines of inquiry, and their pale second-rate imitators. Through this mechanism science at scale reinforces (and in fact, under sufficient political capture imposes) consensus, further annihilating the possibility of the necessary revolutionary synthesis of ideas.

Popular architectures deal with heavy traffic loads by adding logic to cache popular views and resources.

WhyGeneral infrastructure simply takes time to build. You have to carefully design interfaces, write documentation and tests, and make sure that your systems will handle load. All of that is rival with experimentation, and not just because it takes time to build: it also makes the system much more rigid.Once you have lots of users with lots of use cases, it’s more difficult to change anything or to pursue radical experiments. You’ve got to make sure you don’t break things for people or else carefully communicate and manage change.Those same varied users simply consume a great deal of time day-to-day: a fault which occurs for 1% of people will present no real problem in a small prototype, but it’ll be high-priority when you have 100k users.Once this playbook becomes the primary goal, your incentives change: your goal will naturally become making the graphs go up, rather than answering fundamental questions about your system.

This is obviously a powerful playbook, but it should be deployed with careful timing because it tends to freeze the conceptual architecture of the system.

“scaling out is the only cost-effective thing”, but plenty of successful companies managed to scale up with a handful of large machines or VMs

Scaling is hard if you try do it yourself, so absolutely don’t try do it yourself. Use vendor provided, cloud abstractions like Google App Engine, Azure Web Apps or AWS Lambda with autoscaling support enabled if you can possibly avoid it.

I would like to make an appeal to core developers: all design decisions involving involuntary session creation MUST be made with a great caution. In case of a high-load project, avoiding to create a session for non-authenticated users is a vital strategy with a critical influence on application performance. It doesn't really make a big difference, whether you use a database backend, or Redis, or whatever else; eventually, your load would be high enough, and scaling further would not help anymore, so that either network access to the session backend or its “INSERT” performance would become a bottleneck. In my case, it's an application with 20-25 ms response time under a 20000-30000 RPM load. Having to create a session for an each session-less request would be critical enough to decide not to upgrade Django, or to fork and rewrite the corresponding components.

how the features are all on the same relative scale. The relative spaces between each feature’s values have been maintained.

However, this doesn’t mean that Min-Max scaling is not useful at all! A popular application is image processing, where pixel intensities have to be normalized to fit within a certain range (i.e., 0 to 255 for the RGB color range). Also, typical neural network algorithm require data that on a 0-1 scale.

To solve the problem of ‘scaling up’ requires ‘scaling in’ –by this we mean developing the designs and infrastructure needed to support effective use of an innovation.

As student numbers have increased, teaching has regressed for a variety of reasons to a greater focus on information transmission and less focus on questioning, exploration of ideas, presentation of alternative viewpoints, and the development of critical or original thinking. Yet these are the very skills needed by students in a knowledge-based society.