Fable and the end of the free lunch
As frontier AI models like Anthropic’s Fable and OpenAI’s Sol get smarter but more expensive and heavily guarded, many developers are gravitating toward cheaper, “good-enough” alternatives such as DeepSeek, GLM, and other open or low-cost models. Commenters debate whether we’re nearing a plateau in AI capabilities—where gains come mainly from cost and efficiency rather than raw intelligence—or still in a phase of rapid improvement, with future value shifting to specialized models, better tooling, and harnesses that automatically choose the right model for each task. The thread also highlights practical frictions: overzealous safety systems, verbose and cognitively heavy outputs, and uncertainty over who will ultimately capture economic value in a market where powerful models quickly commoditize.
Cost vs capability and “good enough” intelligence
- Many see the real shift in cheap, “good-enough” models (Deepseek v4 Flash, GPT 5.6 Luna, muse/mimo, GLM, Qwen, etc.) that are fast and inexpensive, sufficient for rote coding and everyday assistance.
- Others argue the last 10–20% of capability from frontier models (Fable, Sol, Opus-tier) is disproportionately valuable: better architecture, fewer bugs, higher autonomy, and more strategic insight.
- Some users mix tiers: use top models for planning, design, review; hand off implementation or examples to cheaper models. There’s disagreement on whether this actually saves cost once context and review overhead are included.
Are we hitting an S-curve?
- One camp thinks LLM intelligence is reaching diminishing returns; newer frontier models feel only marginally smarter than earlier Opus/Claude generations, and benchmarks look “gamed.”
- Others counter that capabilities are still improving fast (e.g., math, reasoning, voice), and that claims of plateau have been repeatedly wrong. Evidence for leveling off is called “unclear” by some.
AGI, human brains, and economic impact
- Debate over whether scaling LLMs plus better training could reach “superhuman” or AGI-level intelligence, versus needing brain-like or different architectures.
- Some argue LLMs already exhibit “polyhuman” or domain-superhuman abilities (math proofs, code analysis); critics reply they lack true understanding and online learning, and fail at basics like timezones.
- Disagreement on who “wins” economically: shareholders of first AGI labs, chip/ASIC makers, robotics manufacturers, or society via cheaper services. Some note high competition and likely commoditization.
Guardrails, censorship, and usability
- Strong frustration with Fable/Opus safeguards: frequent self-triggered security checks, downgrades, and refusals (including benign genetics, cooking, or image identification).
- US-aligned models are seen by some as more restrictive than Chinese open models; others warn powerful hardware may become hyperscaler-only, limiting self-hosted freedom.
Hardware, Moore’s law, and specialized silicon
- Disagreement over the Moore’s law analogy: transistor counts kept rising, but single-thread performance stalled, forcing efficiency and parallelism—seen as a parallel to LLM scaling vs training cost.
- Many expect cheaper/faster via custom silicon (Cerebras, Etched, ASICs, dataflow chips, 1‑bit models) and eventual million-tokens-per-second mid-size models; others doubt dramatic cheapening and cite serving costs.
Developer workflows and real-world use
- Heavy use of LLMs as pair programmers, spec reviewers, and voice assistants; some say Opus‑level forever would already be “enough” for their lives.
- Others report Fable is still “stupid” without tight supervision, overly verbose, slow, and prone to self-verification loops, making cheaper or rival models more attractive.