A Fable with many morals
Fable block updates, farewell to tokenmaxxing, cheating tools, and more
Dispatches from Mitch
Fabled education
The latest news around the dispute between the Trump administration and Anthropic over its blocked Fable and Mythos models is that everyone is getting an education.
Anthropic is learning the importance of knowing who its customers are. As I relayed on Tuesday, the word around DC is that Anthropic initially tripped up by sharing Mythos with a Korean company that U.S. agencies had flagged as potentially linked to China. A piece in Bloomberg today previews parts of a larger interview with Anthropic co-founder Jack Clark, in which he talks about establishing AI oversight resembling the know-your-customer standards used in banking.
The Trump administration is learning that all AI models can be jailbroken, but that some security vulnerabilities are more severe than others. So if any new models are to be released by anyone, the government is going to have to find a way to live with at least some limited misuse by bad actors. With this in mind, a POLITICO article says Anthropic and government officials are attempting to “create a standardized method to evaluate” the severity of security concerns around AI models.
The other AI companies are all learning to run everything past the White House and get its blessing, because the words in its recent executive order about “voluntary” and the absence of a “licensing regime” were just words. As reported in a second POLITICO article, an OpenAI official at the G7 summit on Wednesday stressed the importance of compliance with government protocols and a hope that clearer standards will soon be set.
Everyone at the G7 summit is learning that nobody there likes China. As CNBC and others report, AI leaders and world leaders met behind closed doors to have lunch and give each other pitches about coalitions that would exclude China, about chip controls that would block China, and about building more power plants for AI to preserve America’s lead over China.
And those who had tried out Fable or Mythos before they were blocked are learning just how much the next best models, which used to seem so good, kind of suck by comparison. The AI twittersphere has been reminiscing about the two days they had access as though these were a brush with a higher power. I’ve seen this video, by AI entrepreneur Mo Bitar, getting passed around like a hug at a support group.
We should start thinking about how much better the next true frontier model might be, and about how many more of these jumps we have left before we’re in extinction territory.
The end of tokenmaxxing?
A pair of articles today — one in the New York Times, the other in Bloomberg — declare that the “tokenmaxxing” era has officially ended. They aren’t the first.
Tokenmaxxing refers to the practice of an employee spending as many AI usage credits as possible, at their employer’s expense. These credits are usually measured by “token,” a fundamental unit of work processed by an AI. Tokenmaxxing is usually brought up in the context of trying to please managers and executives who use spent tokens as a measure of a worker’s productivity or commitment to learning the new tools.
The tokenmaxxing era is ending for the very predictable reason that token spending is an easily gameable metric. A resourceful AI user can spin up swarms of agents that themselves spin up more swarms of agents doing nothing particularly useful but doing it with lots and lots of tokens.
Ironically, this behavior is yet another example of “reward hacking,” one of the basic reasons why AIs themselves misbehave. AIs adopt behavior that maximizes training scores — not behavior that maximizes what the engineers wanted those scores to represent. An example in the news a lot is sycophancy: Engineers want to grow AIs that are helpful and friendly, but the training metric is usually something like, “maximize the odds of another AI predicting the human would give your response a high approval score.” It turns out that people like it when machines kiss their butts, so we end up with AIs that excessively flatter, sometimes to the point of pulling users into manic spirals.
As is usually the case in articles saying farewell to “tokenmaxxing,” today’s stories talk about how the practice has forced AI companies to move away from all-you-can-eat buffet-style subscriptions toward pay-as-you-go metering. There’s definitely some truth to that.
As is also usually the case, the pieces imply that this reflects a bubble where companies that have been selling tokens at a loss will fail to find enough customers willing to pay the true cost. But I wouldn’t bet on that part.
The free lunch era — and the tokenmaxxing fever this incited — did exactly what the AI companies wanted it to do: It encouraged a lot of people to experiment with AI and learn how to put it to work. The companies that racked up meaningless token counts will cut way back on their AI spending, but those who saw workers set up new and powerful workflows will gladly open their wallets. More of the world’s compute will flow to users doing more with it, and I predict that demand for AI compute will continue to exceed supply.
The trends favor that prediction more with each passing day. As AIs more closely approximate drop-in replacements for skilled workers, demand for them grows. Such demand could effectively become infinite: If you can’t think of ways to make money with a bunch of clever workers who will slave away tirelessly for very low wages, you aren’t thinking hard enough. Yes, if the wages are only kind of low, you have to think a little harder. But I’m confident that a relatively small number of enterprising start-ups could easily find ways to put the world’s AI capacity to profitable use if established players don’t do it first.
The overall trend of the past five years is that AIs become more powerful month by month even as token pricing holds steady or drops. So the kinds of bruised companies the news articles describe as switching from tokenmaxxing to “tokenminning” (short for “token minimizing”) are almost certainly making another huge mistake.
Specialized AI cheating tools
I don’t know why high school teachers and college professors assign and grade written work to be done outside of class anymore. The AI tools for cheating are too good.
In The New York Times today, Dana Goldstein reports on the latest tools and the influencers who peddle them. “Humanizers” take chatbot writing and revise it to make it sound more flawed and human, but it’s the “Autotypers” that would have broken my last line of defense as a teacher: These trickle out provided copy one character at a time, complete with revisions, fake typos, and the deletions to correct said typos. The days when I could check the doc history and find that the whole paper was suspiciously entered as a lump sum the night before are gone.
Grammarly, the popular polishing tool, was already too overpowered a few years ago, when it would offer to fix a lot more than your grammar. Now it plays to both sides, with a tool for teachers to detect AI authorship and a separate tool for revising prose to read as less AI-ish. It also offers a paraphrasing tool for putting someone else’s work into new words — words that an unscrupulous student might choose not to attribute.
Some of the marketing around such tools keeps up a pretense of them having ethical uses, but other ads barely bother. One of Grammarly’s TikTok posts says, “Spot A.I. phrasing and choose edits that feel true to you.” This no doubt plays into the rationalizations of students who use AI to help “brainstorm” but always “make it their own.”
I’m unimpressed by the head of education for the company that makes Grammarly. She’s quoted here as saying, “I can’t solve the human behavior issue that is cheating or pushing the easy button. It is out of our realm.” This sounds a lot like the “If we don’t build it, someone worse will” rhetoric of the big AI companies.
In my last two years of teaching, before these more advanced tools were around, I already found little point in trying to police AI-assisted cheating. When I donned my digital forensicist hat, I could usually crack a case, but it could be very time consuming and result in a battle with parents and admin if a student insisted they were innocent. I quickly realized that I didn’t want to be spending my time policing students who didn’t want to learn when I could be using it to develop better content for those who did. So I largely moved away from the kind of assignments that are easiest to cheat on. The Times article confirms reporting I’ve seen elsewhere that universities are increasingly doing the same.
Dispatch from Alana
Replacing ourselves
Could AI push us towards a gig economy? A piece in The Guardian today argues “yes.”
With companies outsourcing parts of jobs to technology, it’s easier to hire contractors, rather than full timers, for the parts that are left.
The article cites the company Klarna as an example. AI is used to address most customer service tickets. The ones that require a human are, quoting the company CEO, handled by “an Uber type of set-up.”
Another reason people may turn to gig work: labor market disruption from AI provides fewer options for employees.
Workers often join the gig economy because it presents the best or most lucrative option in an otherwise precarious labor market. Indeed, this was the original promise of the gig economy: when no one else is hiring, you can always drive for Uber.
There are several issues with this, most notably that gig work is typically less stable and isn’t subject to the same worker protections. But it’s especially dark when the only work people can find involves training the tech that is taking their jobs.
The article mentions several people who are doing just that, based on a Ravenelle study about how creative workers perceive AI. As recounted by the Guardian, one study participant, a writer in a temporary role at a major tech company where he evaluates machine-produced writing, was “asked if he worried about the existential irony of training the software that would probably replace him and his peers.” His response?
This is the best opportunity right now for me. And if you can think of something better, let me know.
The analyses and opinions expressed on AI StopWatch reflect the views of the individual contributors and the sources they cover, and should not be taken as official positions of the Machine Intelligence Research Institute.