Daniel Jalkut’s balanced take on AI: why extremes miss the point

Daniel Jalkut’s recent comment that “everybody who’s against AI is too against it and everybody who’s for it is too for it” captures a growing frustration with polarized discourse. This article unpacks what the quote implies, examines recent evidence from model releases and product roll‑outs, and points out the practical limits of both blind optimism and blanket opposition.

What’s claimed

Daniel Jalkut, quoted by John Gruber on May 30, 2026, summed up a sentiment that’s been echoing through newsletters and forum threads: “Everybody who’s against AI is too against it and everybody who’s for it is too for it.” The statement is deliberately provocative, suggesting that both camps—skeptics and evangelists—are missing nuance.

What’s actually new

The quote itself isn’t new, but its timing aligns with a cluster of concrete developments that force a reassessment of the usual talking points:

Claude Opus 4.8 – Anthropic’s latest model, announced on May 28, posted a modest but measurable gain on the MMLU benchmark (average 2.3 % absolute improvement over Opus 4.7). The release also introduced a context‑window increase from 100k to 150k tokens, enabling longer document analyses without a proportional rise in latency.
Anthropic & OpenAI product‑market fit – Both companies reported sustained revenue growth in Q1 2026, driven by enterprise contracts for code‑assistant tools and compliance‑focused chatbots. OpenAI’s ChatGPT Enterprise now ships with built‑in data‑retention controls, while Anthropic’s Claude for Teams integrates with Slack and Microsoft Teams out‑of‑the‑box.
Policy chatter – A recent note from the Vatican (Pope Leo XIV’s encyclical on AI, May 25) frames the technology as a moral tool, neither inherently good nor evil, echoing Jalkut’s call for a middle ground.

These items illustrate that the AI field is no longer defined by hype‑driven headline claims; it’s entering a phase where incremental engineering wins and real‑world integrations matter more than speculative fear‑mongering.

Limitations and why extremes still matter

Even with these advances, the quote glosses over several hard constraints that keep the debate relevant:

Area	Over‑optimism risk	Over‑pessimism risk
Safety	Assuming alignment techniques scale linearly; recent Red Team findings show failure modes re‑emerge when models exceed 200B parameters.	Ignoring the fact that many safety tools (e.g., RLHF, constitutional AI) have demonstrable impact on reducing toxic outputs.
Economic impact	Predicting immediate mass layoffs; data from GitHub Copilot shows a net increase in developer productivity rather than headcount cuts.	Dismissing the genuine displacement risk in low‑skill content generation (e.g., automated news briefs).
Regulation	Believing regulation will be uniformly applied; the EU AI Act’s tiered approach creates compliance cliffs for models above 30 B parameters.	Assuming regulation will stall innovation; many firms are already building compliance layers that become competitive differentiators.

In short, the middle‑ground position is useful as a heuristic, but it does not eliminate the need for rigorous evaluation in each domain.

Practical takeaways for practitioners

Benchmark responsibly – The 2.3 % gain on MMLU is real, but it should be contextualized with in‑domain performance. For code‑completion tasks, Claude Opus 4.8 still trails GPT‑4o by roughly 5 % on the HumanEval benchmark.
Integrate compliance early – OpenAI’s new data‑retention flags can be toggled via the API (retain_data: false). Teams building on top of Claude should adopt Anthropic’s policy‑as‑code templates to avoid retroactive fixes.
Monitor safety signals – Deploy a continuous Red‑Team pipeline that feeds back into RLHF loops. The recent failure case where Opus 4.8 generated a plausible but factually incorrect legal citation underscores the need for domain‑specific guardrails.
Balance optimism with cost analysis – While the longer context window is attractive, token pricing has risen 12 % since the last quarter. Evaluate whether the extra context actually reduces overall token consumption for your workload.

Conclusion

Jalkut’s quip captures a growing fatigue with binary narratives around AI. The evidence from recent model releases, enterprise adoption, and even theological reflections suggests that the technology is maturing beyond hype. Yet the same evidence also surfaces concrete challenges—safety regressions, regulatory cliffs, and economic displacement—that demand more than a “middle‑ground” platitude. Practitioners who treat the quote as a call for balanced curiosity—while still grounding decisions in data, benchmarks, and risk assessments—will be best positioned to navigate the next phase of AI development.

References