a16zFormer Microsoft Executive Explains Where We Are in the AI Cycle w/ Anish Acharya & Steven Sinofsky
CHAPTERS
How early are we? The “64K IBM PC” moment for modern AI
Steven frames today’s AI as analogous to the earliest microcomputer era: exciting, but constrained and still poorly understood. The chapter sets expectations that many headline use cases are premature because the tech still fails at basic reliability tasks.
- •AI is compared to the 64K IBM PC era: foundational limits still dominate progress
- •People predict AI will replace core tools (Search, Excel), but capability gaps remain
- •Errors and weak arithmetic/reliability signal an immature platform stage
- •Early-stage energy goes into solving basic “make it work” problems
Lessons from Karpathy: new tools require relearning the human–machine relationship
The group reacts to Andrej Karpathy’s “jagged intelligence” framing and the idea that LLMs invert our typical relationship with computing tools. Productivity will come from learning how to work with these systems’ strengths and sharp edges, not by treating them like traditional software.
- •Karpathy’s metaphors help locate where we are in the AI cycle
- •LLMs as “people spirits” and jagged intelligence change how tools should be used
- •Users must relearn interaction patterns before expecting sustained productivity gains
- •Constraints and trade-offs should guide builders more than hype
Why early platforms obsess over tooling—and why devs lead adoption
Steven argues that early platform transitions are defined by tool-building, and developers are uniquely motivated to force tools into usefulness. Coding becomes a natural early win because the platform’s earliest “customers” are developers themselves.
- •Early stages of any platform are dominated by tool creation and iteration
- •Developers adopt first and adapt tooling to make progress despite rough edges
- •The platform transition is visible and mainstream (news/social), amplifying expectations
- •Tooling maturity is often confused with actual readiness for broad end-users
Vibe writing vs. vibe coding: where autonomy works today—and where it doesn’t
They contrast “vibe writing” and “vibe coding,” debating how much autonomy is realistic. Anish argues writing can reach full autonomy sooner, while Steven stresses that real-world stakes (grades, salary, legal risk) demand verification even when output looks good.
- •Vibe writing feels immediately useful; vibe coding is constrained by production realities
- •Autonomy depends on the cost of being wrong, not just output fluency
- •Human-as-editor emerges as the practical model for many writing workflows
- •Examples of failure modes: hallucinated citations, plausible but incorrect prose
Agents and automation: the slider between no autonomy and full autonomy
Building on Karpathy’s “Iron Man” slider analogy, they discuss the reality that agents won’t arrive as a single breakthrough year. Steven argues we’re in a “decade of agents,” because automation repeatedly hits hard limits around edge cases, trust, and verification.
- •Autonomy should be thought of as a controllable spectrum (no/partial/full)
- •“Year of agents” is dismissed in favor of a longer, messier adoption curve
- •Automation timelines are often overly aggressive compared to historical precedent
- •Verification and exception handling are the hidden costs of agentic systems
Where automation lands first: high-friction, low-judgment tasks (and why incentives matter)
Anish proposes a 2x2: automation arrives first where tasks are high friction but require low judgment (e.g., refinancing shopping). Steven adds an economic layer: markets need differentiation and incentives, so a pure “headless API” future doesn’t align with how businesses attract and monetize customers.
- •High-friction, low-judgment workflows are prime candidates for delegation to agents
- •High-judgment domains (taxes) are risk-laden and harder to automate end-to-end
- •Consumers still want choice and constraints (time, brand, loyalty, comfort)
- •Producers need differentiation/marketing channels; fully commoditized “faceless” offerings break incentives
Correctness vs. judgment: why some domains go fully autonomous and others won’t
Steven distinguishes between domains with formal correctness (chess/Go) and those dominated by uncertainty and human judgment (medicine, taxes, operations). In many real jobs, the work is largely exception handling, which resists clean automation even as tools improve.
- •Formal correctness enables a path to full autonomy; judgment-heavy domains do not
- •Medicine example: uncertainty is intrinsic; “pretend certainty” tools can mislead
- •Radiology shows adoption as augmentation—like better scanners/software, not replacement
- •Taxes illustrate exception trees: automation requires knowing every edge case
The future of product management: ambiguity as the job that doesn’t go away
They address recurring claims that AI will eliminate product management. Anish argues PM work is fundamentally about resolving ambiguity inside complex adaptive systems, so the role may evolve but won’t disappear because organizations will still need judgment and alignment.
- •PM is framed as the discipline of addressing ambiguity that blocks progress
- •Resentment toward PMs is noted, but the core function remains necessary
- •AI may change execution details, but not the need for human decision-making
- •Companies remain complex systems requiring coordination, trade-offs, and judgment
Vibe coding for clout—and the hidden reality that prompts become programming
Steven critiques social-media-driven demos and “it worked instantly” narratives, arguing they often mask long debugging sessions and fragile outputs. He predicts text-to-app will trend toward structured prompting—effectively creating new programming languages and abstractions.
- •Public platform transition encourages performative demos and overclaiming
- •Many ‘easy’ successes hide extensive iteration and manual fixing
- •Structured prompting trends toward a new programming language layer
- •Shipping software still requires robustness, security, and maintainability beyond demos
Programming language hype cycles: from OO and low-code to today’s AI claims
They compare current AI coding hype to earlier waves (object-oriented, Smalltalk, C++, Delphi/PowerBuilder, low-code). Steven’s thesis: transitions routinely overpromise “making programming easy,” delivering mostly constant-factor gains rather than order-of-magnitude improvements.
- •Historical hype: OO promised revolution but took ~10 years to peak and normalize
- •Many languages/tools delivered incremental productivity gains, not 10–100x shifts
- •Low-code’s fade illustrates recurring ‘easy app building’ narratives
- •Today’s ‘no more programmers’ claims mirror past extremes in both directions
Why writing may be the first true order-of-magnitude shift (even with new error types)
Steven argues AI changes the economics of writing more dramatically than past tooling changes, even if accuracy isn’t perfect. They liken it to autocorrect: it removes common friction but introduces new classes of mistakes—still net-positive for many business contexts.
- •Writing appears to be shifting by an order of magnitude in speed/effort
- •Business writing is often already imperfect, making ‘good enough’ highly valuable
- •AI introduces different failure modes (plausible nonsense) rather than eliminating errors
- •The value trade-off resembles early word processors: editing flexibility beats fidelity
AI in creative writing: bestsellers, ‘slop,’ and raising the ceiling for artists
They predict AI-assisted novels will succeed commercially, likely via new creators and pseudonyms, with disclosure arriving later. Anish highlights the challenge of pushing models to cultural edges rather than averages, and both note that most content is middling—where AI can have outsized impact.
- •Expectation: commercially successful novels will be heavily AI-assisted soon
- •Art tension: models average the past; great art often lives at the edge
- •Current surge in low-barrier ‘slop’ increases output but doesn’t define the ceiling
- •AI may unlock new creative workflows once artists become tool-native
Google’s position after I/O: ‘demise’ is overstated, but influence can shift
They reject the idea that Google is “dying,” emphasizing that big incumbents can mount shock-and-awe launches across the stack. The real question is whether Google can change product culture and go-to-market behavior to fit the new platform era, not whether it can demo features.
- •“Demise of Google” is called an absurd framing; large companies rarely vanish overnight
- •Incumbents can deploy ‘shock and awe’—broad releases across many categories
- •Risk is loss of influence during platform transitions, not immediate collapse
- •Key test: can Google transform how it builds and sells products, beyond integrating into search/ads?
