Synthesized from a Grok DeepSearch conversation, April 2026.
| Aspect | 1980s Expert Systems | Modern LLM/GenAI (2022–2026) |
|---|---|---|
| Core technology | Rule-based expert systems on specialized Lisp hardware | Transformer-based LLMs; multimodal and reasoning models |
| Main promise | Machines replicating expert decision-making in narrow domains | General-purpose intelligence; path to AGI |
| Key strengths delivered | Worked well in very narrow, well-defined domains | Real gains in coding, content generation, multimodal tasks |
| Core limitations | Brittle outside scope; expensive to maintain; no learning | Hallucinations; weak causal reasoning; high compute costs; probabilistic |
| Infrastructure bet | Expensive specialized Lisp hardware | Massive GPU data centers; hundreds of billions committed |
| Economic signals | ~$1B+ corporate spend at peak; then sudden collapse | $350B+ in data centers projected; growing ROI questions |
| Disillusionment signs | 1987: hardware market crashed overnight; "AI" became toxic | 2026: Gartner trough entry; 40%+ of agentic projects predicted to fail by 2027 |
| Winter risk | High — led to full winter; research stagnated for years | Moderate (~25–40%) — not a full winter yet, but trough likely |
Expectation inflation. Both eras saw rapid adoption and breathless claims that AI would transform work at scale. In the 1980s, nearly two-thirds of Fortune 500 companies tried expert systems. Today, most large organizations have AI pilots — yet many struggle with production scaling and measurable business value.
Brittleness vs. probabilistic nature. Expert systems were rigid and failed outside narrow scopes. Modern LLMs are flexible but unreliable in high-stakes or novel scenarios — hallucinations echo the "grotesque mistakes on unusual inputs" that doomed expert systems.
Infrastructure overcommitment. The Lisp machine bubble burst when cheaper general-purpose hardware caught up. Today, massive bets on giant data centers face similar risks if efficiency gains (smaller models, better architectures) reduce the need for ever-larger scale.
Funding and hype backlash. Both followed a classic Gartner Hype Cycle. The 1980s version ended in funding cuts and reputational damage. In 2026 the early trough signs are visible: focus shifting to practical deployment, governance, and ROI measurement rather than flashy demos.
Underlying capability is more real. Expert systems were fundamentally limited by rules and couldn't scale with data. LLMs are statistical learners that improve with scale and deliver genuine utility today. Progress feels more incremental and grounded.
Data and compute foundation. Modern AI benefits from internet-scale data and hardware advances. No equivalent "general workstation" has fully obsoleted the need for specialized AI infrastructure yet, though smaller and open-source models are gaining traction.
The field has diversified. A calibrated, multi-model approach — deep work, research, real-time, multimodal covered by different tools — avoids single-point dependency and treats AI as leverage rather than magic. Veterans of the 1980s winter often advocate exactly this.
Having been at Teknowledge in 1985 — at the peak of the boom, just before the winter hit — is not an abstraction. That era instilled a deep preference for inspectable, deterministic, spec-first systems: YAML as single source of truth, Action–State patterns that work for humans and LLMs alike, reproducible environments via Mise and Brewfile.
The pragmatic response to today's hype: upgrade AI tools when they solve real workflow friction (Claude Max 5x eliminated afternoon quota exhaustion), diversify across models so no single limitation becomes a blocker, and build systems with deterministic runtimes that don't depend on the probabilistic layer being reliable.
In short: the 1980s winter was a harsh teacher about overhype, maintenance costs, and infrastructure fragility. Modern hype shares those DNA markers but rides on more capable foundations. We're likely in a trough/consolidation phase in 2026 rather than an imminent full winter — but the risk remains if ROI doesn't materialize broadly.