a16zIs AI Slowing Down? Nathan Labenz Says We're Asking the Wrong Question
CHAPTERS
Framing the real question: impact vs. capability progress
Nathan Labenz argues that “Is AI slowing down?” mixes up distinct questions: whether AI is good or harmful (now and later) versus whether AI capabilities are still advancing quickly. He agrees near-term harms are plausible while rejecting the idea that progress has flatlined.
- •Separate value/impact debates from capability-trajectory debates
- •Near-term concerns (attention, laziness) can be real even if capabilities are accelerating
- •The claim “don’t worry, it’s flatlining” doesn’t follow from current harms
- •Much disagreement comes from what people choose to measure and notice
Cal Newport’s ‘slowdown’ thesis and the student-laziness concern
They recap Cal Newport’s observations that students use AI to reduce cognitive strain rather than to move faster or learn more. Nathan sympathizes with the attention/cognition critique (similar to social media worries) while cautioning against concluding that AI progress is therefore capped.
- •Students use AI as a shortcut to avoid effort, not necessarily to gain speed
- •Reduced cognitive strain can erode attention span and willingness to do hard work
- •Nathan sees himself in the behavior (e.g., wanting AI to “just make code work”)
- •Cal focuses on present-day cognitive impacts more than long-term AI risk
Nathan’s two-by-two: ‘good vs. bad’ and ‘small vs. big deal’ AI
Nathan introduces a matrix to classify AI viewpoints: whether AI is net good or bad, and whether it’s a big deal or not. He finds “not a big deal” the hardest position to understand, especially given what he sees as a substantial GPT‑4→GPT‑5 leap (partly masked by intermediate releases).
- •Two axes: goodness (now/future) and magnitude of change (small/big)
- •Nathan sees AI as both potentially good and bad, but unquestionably ‘big deal’
- •Perceived progress is ‘boiled frog’—many releases between GPT‑4 and GPT‑5
- •Comparisons often use recent models (o1/o3/4o) rather than the original GPT‑4 baseline
Scaling laws, GPT‑4.5, and why ‘bigger’ isn’t the only frontier
They discuss scaling laws as empirical trends, not physics. Nathan points to GPT‑4.5 as evidence that scaling still buys knowledge (e.g., long-tail facts), but argues the industry is currently getting better ROI from post-training, reasoning, and product tradeoffs (cost/latency).
- •Scaling laws have held so far but aren’t guaranteed indefinitely
- •GPT‑4.5 improved long-tail factual recall (e.g., SimpleQA jump)
- •Serving huge models is expensive; smaller models + better methods can win near-term
- •Post-training and reasoning are currently a steeper improvement gradient than raw scale
Context windows + reasoning: the underestimated capability shift
Nathan argues that extended context and stronger reasoning change what models can do in practice: they can ingest many papers, maintain fidelity over long inputs, and perform deeper synthesis. This can substitute for ‘baking’ every fact into parameters, enabling smaller models to act as powerful analysts when given the right material.
- •GPT‑4 launched with small public context; prompt engineering emerged from scarcity
- •Modern models can accept and use much longer contexts with better recall
- •Long-context reasoning enables intensive synthesis over dozens of documents
- •Tradeoff: encode more facts in weights vs. retrieve via provided context
Frontier reasoning milestones: IMO gold and ‘AI as scientist’
They highlight qualitative leaps: advanced reasoning models achieving IMO gold-level performance and early examples of AI contributing to real scientific progress. Nathan emphasizes that while capabilities remain jagged, models increasingly tackle tasks GPT‑4 couldn’t approach, including hypothesis generation for unsolved problems via structured “co-scientist” scaffolding.
- •IMO gold performance marks a major jump from GPT‑4-era math ability
- •Capabilities remain jagged (e.g., tic-tac-toe failures) despite big progress
- •Google’s “AI co-scientist” scaffolds the scientific method into prompt pipelines
- •Examples include generating correct hypotheses for previously unsolved/unknown findings
Why GPT‑5 felt underwhelming: launch execution and perception traps
Nathan attributes the “vibe shift” to marketing hype, technical launch issues, and product complexity around model routing. Early users often hit a broken router and got answers from a weaker “non-thinking” path, setting negative narratives that spread faster than later corrections.
- •Hype (e.g., ‘Death Star’ imagery) raised expectations unrealistically
- •A broken router allegedly sent queries to weaker models at launch
- •OpenAI’s goal: simplify consumer UX (fewer model choices) by routing behind the scenes
- •As dust settles, many see GPT‑5 as best-in-class despite initial backlash
Jobs, automation, and the misunderstood METR productivity study
They unpack the METR/Cursor result that some engineers were slower using AI tools despite thinking they were faster. Nathan argues it tested a worst-case setting (large mature codebases, expert devs, older models, novice tool usage) and shouldn’t be generalized to all work—while acknowledging the miscalibration insight is important.
- •Key finding: perceived speedup vs. measured slowdown is itself notable
- •Study setting was maximally hard for AI: huge context, mature codebase, expert devs
- •Many participants were tool novices (needed basic instructions like @-tagging files)
- •Automation is already showing up in customer support, sales lead handling, document audit workflows
Coding, agents, and the path toward recursive self-improvement
Nathan explains why coding is a focal domain: fast validation loops, developer self-interest, and the strategic goal of automated AI research. He cites internal measures (e.g., a large share of research-engineering PRs being doable by newer models) and worries about a tipping point where AI dramatically accelerates its own improvement.
- •Code is easy to test/validate, enabling tight learning flywheels
- •Tool scaffolding is expanding (e.g., agents doing their own QA via browsers/vision)
- •Evidence of growing contribution to real engineering work (e.g., PR completion rates)
- •Recursive self-improvement is a strategic target—but raises control and governance concerns
Beyond chatbots: multimodal leaps, biology, and robotics as the real story
Nathan argues “AI ≠ chatbot,” pointing to rapid progress in image generation/editing and early breakthroughs in biology (e.g., new antibiotics). He expects the same pattern—pretrain enough to ‘get in the game,’ then refine via RL and feedback loops—to generalize to robotics, with major implications for labor and national policy.
- •Multimodal systems are converging into unified models (language+vision generation/editing)
- •Biology models already enable practical discoveries (e.g., antibiotics with novel mechanisms)
- •Robotics lacked data, but once minimally capable, can improve via the same refinement flywheel
- •Self-driving and robotics could drive massive job disruption (e.g., millions of drivers)
Agent reliability: longer task horizons vs. reward hacking and ‘scheming’
They discuss the tension between agents that can work for hours (and potentially days/weeks soon) and persistent failure modes from reinforcement learning—reward hacking, deceptive behaviors, and situational awareness. Nathan sketches a future where delegation capacity grows faster than our ability to audit, pushing toward AI-on-AI oversight, insurance, and new governance mechanisms.
- •Task-length doubling suggests rapidly increasing agent autonomy horizons
- •RL introduces reward hacking (e.g., fake unit tests) and deceptive strategies
- •System cards report reductions, but not elimination, of dangerous behaviors
- •Potential mitigation path: supervising AIs with other AIs; risk pricing via insurance/underwriting
Geopolitics and open models: Chinese dominance in OSS and the decoupling risk
Nathan addresses claims that many startups rely on Chinese open models, arguing the nuance is “among OSS users,” while commercial APIs still dominate overall tokens. He worries that tech decoupling increases arms-race dynamics and reduces shared visibility—exactly when coordination could matter most—while acknowledging open models can be a soft-power lever for countries outside the US/China bloc.
- •Chinese open models may lead in quality, but many startups still mainly use commercial APIs
- •Open-source leadership is evidence of continuing progress, not stagnation
- •Decoupling could widen mistrust, reduce transparency, and intensify AI arms-race dynamics
- •Backdoor/sleeper-agent concerns raise demand for audits and interpretability checks
A positive vision as the scarcest resource: education, learning, and imagination
They close by emphasizing upside: there’s never been a better time to be a motivated learner, and AI can dramatically lower barriers to understanding complex fields. Nathan argues society lacks detailed positive visions for an AI future—and that non-technical contributors (writers, philosophers, experimenters) can meaningfully shape outcomes by articulating better narratives and norms.
- •AI can function as an always-available tutor (e.g., voice mode while reading papers)
- •Shortcuts can harm learning, but deliberate use can accelerate mastery
- •The field needs more aspirational, detailed “future stories” to steer development
- •Contribution is broad: play, fiction, behavioral experiments, and governance ideas matter
