AI Hallucinations, the 'Forever Problem' in AI & Humans.
Three AI stories landed in one news cycle on April 21 — a pre-IPO M&A bet on AI coding, OpenAI’s return to images inside ChatGPT, and a top global law firm formally acknowledging AI hallucinations in its own work product.
Each lands on the same thesis from a different angle: hallucinations are a forever problem — an architectural feature of probabilistic systems meeting deterministic computing, to be worked around domain by domain, not patched out by the labs. Anchor post written over a year ago and holding up: AI-RTZ #741 — AI Hallucinations the ‘Forever Problem’.
Three Key Takes today:
(1) AI Coding Is the Best-product market fit (PMF) AI App So Far — And the Domain Where Hallucinations Hurt Least. “AI Coding is the most financially successful ‘product-market-fit’ AI application to date this early in the AI Tech Wave. It’s even more important than ChatGPT over three years ago. Today, Anthropic Claude Code/Cowork, OpenAI Codex, Cursor’s products, and a host of other coding products are all being pursued hard by the leading AI companies here and in China. Yet, they too have and will have ‘hallucinations’ baked into the how they work. Just less problematic than other domains AI is being applied to. And it’s because the hallucinations here for the most part can be caught during runtime where the software meets the hardware.” The catalyst: NY Times — SpaceX/Cursor deal. Elon’s SpaceX/xAI cut a $60B 12-month call option on Cursor — with a $10B break fee — right before the SpaceX mega-IPO. The pre-IPO framing: AI coding is where the economics already work, so you buy the best standalone before going public. Why coding is the “good” domain: Simon Willison — Hallucinations in code are the least dangerous. The compiler is the ground truth. Tests fail loudly. The cost of a hallucinated function is a red squiggle, not a lawsuit. Compute + fabs backdrop: AI-RTZ #974 — SpaceX prepping for 2026 mega AI IPO and #1034 — Musk boosts SpaceX/xAI pre-IPO, Terafabs and beyond.
(2) OpenAI Is Fusing Images Back Into ChatGPT — Because the Compute Economics Work better Bundled.“OpenAI is fusing in the image capabilities of AI into its core products. Rather than creating new separate services that then need to work on their own financial feet and somehow pay back the relatively higher compute required for AI images and video. For now.” The return: The Verge — OpenAI’s new ChatGPT image engine. Image2 folds back inside ChatGPT, not as a Sora-style standalone. The hallucination tax is still there: Gary Marcus — ChatGPT’s powerful new image engine still hallucinates — priced in, not solved. Why the Sora side-quest got trimmed: AI-RTZ #1037 — OpenAI trims AI applications down. Standalone products have to carry their own compute cost; they couldn’t. What a bundled-image win actually looks like: AI-RTZ #674 — Studying OpenAI’s Global Studio (the Ghibli moment). The virality happened inside ChatGPT. Back to the Future.
(3) Hallucinations Are Architectural — Every Vertical Needs Its Own Operational + Governance Playbook. “AI Hallucinations will be an ongoing issue as probabilistic systems work with deterministic computing systems in hybrid fashion. There will be issues in almost every vertical domain. And every vertical domain will need to figure out their own operational and governance imperatives to work with AI despite the hallucinations. They for now at least, are a ‘Forever Problem’. And feature.” The story: Financial Times — Sullivan & Cromwell AI hallucinations admission. A top-three global law firm formally acknowledging AI hallucinations in its own work product. The through-line across all three Takes: probabilistic meets deterministic. Coding catches the errors cheaply at the compiler. Images absorb them cosmetically behind bundled distribution. Law firms now price them in explicitly via process + disclosure. Different domains, different economics of error — same architectural reality. Plan around it; don’t wait for the labs to fix it.
Plus: Gadget AI — “The issue of trust in ambient AI sensors capturing data and delivering AI reasoning, agentic and operating capabilities, is just at the beginning. The solutions go beyond technical innovation. And have to do with societal and governance norms. They will need to adjust and adapt, with appropriate modifications to the underlying AI technologies. This will take more time and expense than currently being assumed.” The item: Reuters — Meta to start capturing employee mouse + keystrokes for AI training data. Not new — echo of the Microsoft Recall feature in Copilot Windows PCs (AI-RTZ #877). The underlying frame: AI-RTZ #382 — Scaling AI Trust is Job 1. Hallucinations in code + images + legal briefs are the visible trust problem. Silent surface capture is the invisible one. Same trust ledger — and the bill takes longer to come due than tech-industry models assume.
Bonus — today’s AI-RTZ companion #1064 covers “The Velvet Rope AI Strategy” — Anthropic and OpenAI quietly gating advanced features (Anthropic just pulled Claude Code from the $20/month base tier) behind higher-priced subscriptions. The pricing story underneath the hallucination story.
Closing Questions — biggest trust reservations with AI: connecting AI systems to complete personal computing surfaces. Full keystroke / mouse / screen capture is where the trust math breaks — the failure mode isn’t a bad output, it’s involuntary inventory of everything you do. Ties directly to today’s Meta Gadget AI item and the AI-RTZ #382 trust-as-bottleneck frame. Companies I’ve trusted despite reservations: Google Gemini (privacy engineering maturity), Anthropic Claude Cowork (explicit trust/safety posture + request-per-app permissions), Perplexity more selectively (narrower attack surface as an answer engine vs. a full assistant). Stay tuned.
Clips from today’s episode
[Short — AI Hallucinations: A Forever Problem](YouTube Shorts link) Hallucinations aren’t a bug to patch — they’re an architectural feature of probabilistic AI systems. Even Sullivan & Cromwell, a top-three global law firm, just had to admit it. Plan around it; don’t wait for the labs to fix it.
[Short — Musk’s $60B Cursor Play](YouTube Shorts link) Elon’s SpaceX/xAI cut a $60B 12-month call option on Cursor — with a $10B break fee — right before the mega-IPO. AI coding is the best-PMF AI app to date. What this pre-IPO consolidation signals.
About AI Ramblings Daily (ARD), and AI-RTZ
Both are daily. Both are free. Both are about AI. But they’re different mediums carrying different messages.
AI-RTZ is the morning text — a deeper written take on one idea, published by at least 5 AM EST. Today: post #1064.
AI Ramblings Daily is the afternoon video + podcast — my ad hoc takes and perspective on the day’s AI issues & news flow, around 20 minutes, with short 1-2 minute clips for quick topic views. Today: episode #60.
Subscribe to either or both on michaelparekh.substack.com. They run as separate Sections you can opt into or out of.
Links used in today’s show (already embedded inline above; listed here for reference)
Framing — AI Hallucinations as the Forever Problem:
Take 1 — AI Coding is Best-PMF / Musk’s $60B Cursor Play:
-
NY Times — SpaceX/Cursor deal (MP gift link — always embed, never paste raw)
-
Simon Willison — Hallucinations in code are the least dangerous
-
AI-RTZ #1034 — Musk boosts SpaceX/xAI pre-IPO, Terafabs and beyond
Take 2 — OpenAI Fuses Image2 Into ChatGPT:
-
Gary Marcus — ChatGPT’s powerful new image engine still hallucinates
-
AI-RTZ #674 — Studying OpenAI’s Global Studio (Ghibli moment)
Take 3 — Sullivan & Cromwell Hallucination Admission:
-
Financial Times — Sullivan & Cromwell AI hallucinations admission (MP gift link — always embed, never paste raw)
Plus — Gadget AI on Meta data capture:
Bonus — today’s AI-RTZ companion:
-
AI-RTZ #1064 — The Velvet Rope AI Strategy (swap placeholder with public URL after publish)
(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us her
