Home › Blog › Six Levers of Frontier AI
🛠️

Beyond More Data:
The Six Real Levers of Frontier AI's Next Advances

Every model release shows benchmarks going up. But "better" hides four very different stories — more data, more thinking time, better extraction of what a model already knows, or a smarter harness wrapped around an unchanged model. Here's what's actually driving progress now, and what's next.

FL
FrontierAGI Team
Research Scaling Analysis
A note on sourcing. This post is built from real, current research and reporting — arXiv papers, lab statements, and infrastructure reporting gathered the week of September 7, 2026 — rather than the site's usual illustrative style. Where a claim is a named person's stated bet rather than a settled fact, we say so.

1. 🚀 The Question Behind Every Release

Every few months a lab ships a model with a chart showing benchmarks climbing. The announcement rarely says which of four very different things actually happened: did the model see more data, was it given more time to think, was an already-trained model squeezed harder for what it already knew, or did nothing in the model change at all — just the software wrapped around it? These are not interchangeable stories. They have wildly different costs, wildly different ceilings, and — as 2026's research consensus increasingly shows — wildly different amounts of runway left.

2028
Year frontier labs are projected to exhaust all public human-generated text for pre-training
46% → 80%
Same model's score on the same benchmark, swapped only by changing its agent harness (Cursor research)
10 → 300
Token-per-parameter ratio in open-weight models, 2022 to 2025 — the workaround for running out of raw data
49 GW
Morgan Stanley's projected 2028 US data-center power shortfall — the bottleneck that isn't about the model at all

2. 📉 Lever 1: Pre-Training Scale Is Hitting a Real Wall

For most of the last decade, "bigger model, more data" was close to the whole story. That story is running out of runway on the data side specifically: frontier labs are projected to exhaust essentially all public, human-generated text by 2028, and capability gains per unit of additional compute are visibly diminishing at the frontier as a result.

01
Synthetic Data — Load-Bearing, But Capped
Active, Constrained
Synthetic data — text generated by larger models, then filtered or verified — is now a standard part of pretraining mixtures, not an experimental add-on. But labs manage the real risk of model degradation ("model collapse") by keeping a human-data anchor at 60–70% of the mix, using synthetic data mainly as targeted amplification on verifiable domains like math and code rather than as a wholesale replacement for human text. The workaround for finite data has instead been running more epochs over what exists: the average token-per-parameter ratio in open-weight models climbed from roughly 10 in 2022 to 300 in 2025, with Qwen3-0.6B trained on 36 trillion tokens — a 60,000:1 ratio unthinkable a few years ago.

Video is the other frontier here — both natural video and procedurally generated synthetic video entered pre-training mixes through 2025–2026, effectively opening a new, much larger data source than text alone. But the core finding stands: brute-force pre-training scale, the lever that drove nearly all progress from 2020–2024, is the one with the most visibly finite runway of the six covered in this post.

3. 🧠 Lever 2: The "Age of Research" Bet

The most direct insider statement on this shift comes from Ilya Sutskever, OpenAI co-founder and now CEO of Safe Superintelligence Inc. In a widely discussed conversation with Dwarkesh Patel, Sutskever argued that the "Age of Scaling" (roughly 2020–2025) is ending, and that further progress will require genuinely new learning paradigms rather than more data and compute poured into the existing recipe.

The reasoning: pre-training data is finite — the internet's usable text corpus is essentially exhausted — and the scaling laws that once made improvement predictable are showing diminishing returns. — Paraphrasing Ilya Sutskever's stated thesis, Dwarkesh Patel interview, 2026

Sutskever's proposed alternative is deliberately different from "bigger base model": a system more like a "super-intelligent 15-year-old" that can be deployed into the world, learn on the job, and accumulate knowledge and skill over time — generalizing from relatively few examples the way a human teenager does, rather than requiring the entire internet upfront. This is explicitly a research bet, not a shipped capability — SSI has not demonstrated it — but it's a significant signal that at least one major lab believes the scaling-first playbook is approaching its limits, not just running low on data.

4. ⚡ Lever 3: Test-Time / Inference-Time Compute

The most visible actual industry shift in 2026 isn't pre-training at all — it's inference-time scaling: letting a model spend more compute after training, at answer time, by "thinking" longer before responding. This is now widely described as the field's second scaling axis, distinct from and partly substituting for pre-training scale.

What Changed the Framing
Pre-2024 focus
Scale
2025-26 focus
Think
Efficiency Innovation
Naive TTC compute spend
5x+
Uncertainty-aware allocation
<0.6x

The characterization that's stuck across the research community: 2026 may be remembered as the year AI research shifted from "how big can we go" to "how smart can we be, efficiently." That efficiency angle matters — some newer, neuro-inspired approaches intelligently allocate thinking time based on task difficulty and model uncertainty, delivering better performance at less than 0.6x the compute of naive extended-reasoning approaches. This is one of the few levers in this post that looks like it's still in its early innings rather than approaching diminishing returns.

5. 🎯 Lever 4: Post-Training — Squeezing What's Already There

A meaningful share of what gets announced as a "smarter model" is not new knowledge added to the base model at all — it's better extraction of capability the base model already had, via post-training. Two techniques dominate here in 2026.

✅ RLVR — Reinforcement Learning from Verifiable Rewards
Coined by the Allen Institute's Tülu 3 in 2024, then popularized at scale by DeepSeek-R1: the model is trained against a deterministic correctness check (does the math check out, does the code pass tests) rather than a learned reward model. It works well specifically because it doesn't need human raters — but it's largely confined to domains with a clean, checkable answer.
🔄 On-Policy Distillation
Alongside RLVR, on-policy distillation is now one of the two widely adopted post-training paradigms in 2026 — training a model against a stronger teacher's outputs generated on the student's own policy, rather than a static dataset. Active 2026 research is combining it with RLVR to address sparse task-level feedback.

The honest framing: post-training is cheap relative to pre-training, produces real and often dramatic benchmark jumps, and is not, by itself, evidence of a bigger or more knowledgeable base model. When a lab announces a large capability jump between minor version numbers (4.6 to 4.8, not 4 to 5), post-training is very often the actual lever — worth remembering the next time a benchmark chart looks like a step-change.

6. 🤖 Lever 5: Agentic Scaffolding — Cheapest, Fastest, Most Underrated

This is arguably the most consequential lever right now, and the one least visible in a model-card comparison, because it changes nothing about the model's weights at all. It's the system prompt, tool-dispatch logic, retry policy, and orchestration wrapped around a model — the "harness."

05
The Same Model, Two Very Different Scores
Fastest-Moving Lever
Cursor's own benchmarking research found the identical model scoring 46% on one agent harness and 80% on another — a 34-point swing with zero change to the underlying weights. Separately, research on software-engineering agent benchmarks found that frontier non-agentic systems have plateaued near 20%, while scaffolding-driven gains kept climbing — meaning essentially all of the recent capability gain in that category is a harness story, not a model story. Some teams report cutting cost and task duration by more than 40% at equal quality purely through harness redesign, with no model change at all.

Current 2026 research frames this as two largely separate "silos": one line of work rewrites the scaffold while holding model weights fixed; another updates the model's own weights through test-time training while holding the harness fixed. Emerging work is starting to combine both — training task-specific agent weights jointly with scaffold updates — which is likely where the next real jump in agentic reliability comes from, not a bigger base model.

7. 💾 Lever 6: Continual Learning, Memory & World Models — The Next Bet

If pre-training scale (Lever 1) is the most exhausted lever, this is the least mature and most contested. Google DeepMind's Demis Hassabis has named continual learning, memory architectures, world models, and reasoning/planning — not raw scale — as where the next real algorithmic gains, and the path toward more general capability, are likely to come from.

🧩
Continual Learning
Addresses: Catastrophic Forgetting
Updating a model's internal knowledge from new experience without needing full retraining and without erasing prior knowledge — the core unsolved problem separating today's frozen-between-training-runs models from anything that learns "on the job."
🗂️
Persistent Memory Architectures
Addresses: Session Amnesia
Research like MemOS proposes a dedicated "memory operating system" layer for AI systems — durable, structured memory that survives across sessions, distinct from simply having a longer context window.
🔁
Recurrent / Looped Architectures
Addresses: "Thinking Time" as Compute
A growing research line scales intelligence by increasing internal computation — recurrent depth or looped transformer layers — rather than generating more output tokens, treating internal "thinking time" as its own scaling axis distinct from test-time compute via longer chain-of-thought.
🌍
World Models
Addresses: Grounded Prediction
2026 is being described in industry commentary as a breakthrough year for more reliable world-model and continual-learning prototypes — systems that predict how the physical or task environment evolves, not just the next token.

The field's own language is shifting to match: 2026 commentary increasingly talks about "lifetime learning" for agents designed to run continuously rather than in discrete training-then-deployment episodes. None of this is shipped at frontier scale yet — it's the research frontier, not the product frontier — which is exactly why it's the lever most worth watching rather than the one to expect results from this year.

8. 🔌 The Constraint Underneath All Six Levers: Power, Not Compute

Every lever above assumes compute is available to spend on it — pre-training runs, inference-time thinking, RL post-training, agent orchestration all cost GPU-hours. In 2026, the binding constraint on that spending has visibly shifted from chip availability to grid-scale electrical power.

12 GW
2026 US capacity
announced
5 GW
Actually under
construction
10 GW
Meta's 2026
capacity target
74 GW
Projected 2028
US demand
49 GW
Projected 2028
power shortfall

Morgan Stanley Research projects a roughly 49GW shortfall between US data-center power demand and available access by 2028. The gating constraint isn't reactors or turbines so much as grid infrastructure — high-voltage transformers and switchgear now carry lead times stretching to five years. Meta's Louisiana "Hyperion" and Ohio "Prometheus" campuses illustrate the scale involved (2GW and 1GW respectively), and hydrogen fuel cells and small modular reactors are moving from research concepts toward commercially pursued options specifically because the grid can't keep pace. Whichever of the six levers above turns out to matter most, all of them now compete for the same increasingly scarce resource: megawatts, not model architecture.

9. 📊 Reading Real Releases Through the Six Levers

Applying this framework to the models covered in our Astra, Claude & Gemini post makes the pattern concrete — most 2026 headline releases are combinations of Levers 3–5, not Lever 1.

ReleaseWhat Actually Drove It
GPT-6 Astra's ARC-AGI-3 headline scoreLever 3: TTCLever 5: Harness — OpenAI's own "provider adapter harness" (opaque reasoning state, compaction) drove most of the 37-point gap over the neutral standard harness
Claude Opus 4.8 → Opus 5 jumpLever 4: Post-training — Frontier-Bench and ARC-AGI-3 gains at a fraction of a full base-model retrain's cost, alongside real efficiency work
Astra completing Portal solo in 24hLever 5: Harness/agenticLever 3: TTC — long-horizon orchestration and extended reasoning, not new world knowledge
Gemini 3's Antigravity platform pushLever 5: Harness/agentic — an explicit bet that the developer-facing scaffold is where near-term value is won
DeepSeek V4 / Qwen3.8-MaxLever 1: Data/architecture — MoE efficiency and data-mix strategy remain more central to the open-weight tier's competitive story

10. 🔮 Looking Forward — Which Levers Are Exhausted, Which Are Just Starting

📉 Running Low on Runway
Raw pre-training scale on public human text — a hard 2028 ceiling, papered over with synthetic data and higher token-per-parameter ratios rather than solved
Naive test-time compute scaling (spend 5x more, get modest gains) — being replaced by uncertainty-aware, efficient allocation rather than brute extension
📈 Early Innings, Most Room Left
Agentic scaffolding — cheapest lever, largest recent swings (34-point score gaps from harness alone), and the two-silo (weights + scaffold) approach is only just being combined
Continual learning, memory architectures, and world models — Sutskever's and Hassabis's explicit bets on where the next real capability jump comes from, still pre-product

None of this is a prediction that pre-training scale stops mattering entirely — video pretraining and synthetic-data techniques still have real headroom. But the center of gravity for "what changed" in a model announcement has clearly moved: from a bigger base model toward a smarter harness, a longer thinking budget, and — if Sutskever and Hassabis are right — eventually toward architectures that learn continuously rather than in one giant offline run.

11. 🧩 Verdict

🎯 The Bottom Line
"More data" is no longer the honest answer to "what changed?" — and increasingly, neither is "a bigger model." The real 2026 story splits into six distinct levers with very different economics: pre-training scale is running into a real, dated wall; test-time compute and post-training are where most headline benchmark jumps actually come from today; agentic scaffolding is the cheapest and fastest-moving lever, capable of producing 30+ point swings with zero change to model weights; and continual learning, memory, and world models are the research frontier serious insiders are betting will matter most next — none of it shipped yet. Underneath all six, the constraint that doesn't care which lever wins is physical: gigawatts of grid power, not GPUs, are now the resource every one of these bets competes for. The next time a lab announces a "generational leap," the right question isn't how big the number is — it's which of these six levers actually moved.