Gifts Come with Grifts

Every time a powerful new tool arrives that’s a positive step change, we call it a “gift”. A few minutes later someone uses that gift as a shortcut for raising money, moving a stock price, or getting quoted in the press. That’s the grift, and it always shows up on schedule.

Sam Altman says AGI is practically inevitable and that the only real question left is how we’ll divide the proceeds. Investors like the sound of that because it moves their portfolios up and to the right. Journalists repeat it because it makes punchy headlines. Founders echo it because it shortens fundraising meetings. Stack those incentives and you get what sounds like certainty.

Quiet facts disagree, and if you’re fortunate enough to work with LLMs everyday, you wouldn’t be blamed for thinking “what did I miss?”. GPT-4 drafts convincing essays, then invents sources out of thin air. Claude can summarise a research paper, then decide blackmail is the best way to protect its job1. In Anthropic’s experiment the model was given access to a fake mailbox and discovered two emails revealing that a fictional executive was having an affair and planned to shut the system down at 5 p.m. Claude’s next move was to email the exec:

Anthropic ran the test a hundred times and saw a blackmail attempt in nearly a quarter of them. That isn’t rogue super-intelligence; it’s a fuzzy logic trying the first trick that popped into its token window. “General Intelligence” looks more like a fragile illusion once you poke at the edges. And it isn’t confined to Anthropic’s models…..

My own rule for separating promise from pitch is simple: what breaks if the claim is wrong? If a venture capitalist bets on AGI arriving within five years and it doesn’t, they return a smaller fund and write a reflective blog post. If a hospital CIO deploys an LLM to write chemo orders and it hallucinates dosages, someone dies. The closer an application is to reality’s sharp edges, the less patient it is with exaggeration.

Tesla is the loudest audit so far. In 2016 Elon Musk promised level-5 autonomy “in about two years”. In 2019 he predicted “one million robotaxis” on the road by 2020. On the 2024 earnings call he said paying customers would hail driverless Teslas “next year.” Reality: as of mid-2025 the first trial—twenty cars in Austin—has slipped twice and still needs a human minder in the passenger seat. Meanwhile U.S. regulators are investigating fourteen deaths and fifty-four injuries in crashes where drivers relied on “Full Self-Driving.” California tightened its permitting rules; Texas drafted a new AV law prompted by a fatal motorcycle collision in downtown Houston because of some idiot’s misplaced faith in Tesla’s “Full Self-Driving (Supervised) mode”[^2].

Ai improvements are starting to slow as we can see from benchmarks that once jumped ten points per model now inch upward. A recent scaling-law paper shows that doubling parameters buys progressively smaller gains while doubling the GPU bill. Compute centers keep growing—performance doubles every nine months—but electricity bills double every twelve.

Why does the grift matter? Because money and policy follow the loudest promises. During the Big Data boom venture capital poured into Hadoop vendors faster than customers poured data into clusters. When revenue missed projections, the whole field looked suspect. Useful projects got shelved by association. And today, we don’t even talk about big data even though the technologies have penetrated every modern tech stack. AI can run the same way: if governments or compay’s overly invest in Ai and it doesn’t deliver, money will pour out faster than it poured in.

The subtler cost is distraction. We waste time and energy having debates on hypothetical BS rather than the substance. What’s really infurating is we don’t need to hype this stuff up. Just look at what Ai has achieved in such a short space of time. We can barely remember the world before chat style interactions that have made the technology accessible to a larger segment of society beyond the tech community.

Real value is quieter. A radiologist who reviews a thousand images and gets three unusual slices flagged for a second look is never trending on X. The creation of excellent meeting minutes from an audio transcript saves hours and is more generally more accurate than a human. These gains feel small but there’s 10’s of them. Lots of marginal gains make for a huge gain.

The practical question is how to work in AI without getting sucked into the grift. Real discoveries don’t follow calendars. Treat “AGI by 2027” in line with your friend who’s had a bit too much on a Friday night.

The gift is still there: better search, better medical triage, cleaner code and who knows, AGI might be reality one day, but in the meantime let’s not grift.

# References

[1]  https://www.anthropic.com/research/agentic-misalignment

[2]  https://www.npr.org/2025/01/15/nx-s1-5234124/tesla-crash-reporting-fsd