OpenAI’s CEO confirms the corporate isn’t coaching GPT-5 and “gained’t for a while”

In a dialogue about threats posed by AI techniques, Sam Altman, OpenAI’s CEO and co-founder, has confirmed that the corporate just isn’t at present coaching GPT-5, the presumed successor to its AI language mannequin GPT-4, released this March.

Talking at an event at MIT, Altman was requested a few latest open letter circulated among the many tech world that requested that labs like OpenAI pause growth of AI techniques “extra highly effective than GPT-4.” The letter highlighted considerations concerning the security of future techniques however has been criticized by many within the trade, together with a variety of signatories. Consultants disagree concerning the nature of the risk posed by AI (is it existential or extra mundane?) in addition to how the trade may go about “pausing” growth within the first place.

At MIT, Altman mentioned the letter was “lacking most technical nuance about the place we’d like the pause” and famous that an earlier model claimed that OpenAI is at present coaching GPT-5. “We’re not and gained’t for a while,” mentioned Altman. “So in that sense it was kind of foolish.”

Nevertheless, simply because OpenAI just isn’t engaged on GPT-5 doesn’t imply it’s not increasing the capabilities of GPT-4 — or, as Altman was eager to emphasize, contemplating the security implications of such work. “We’re doing different issues on high of GPT-4 that I believe have all kinds of issues of safety which might be vital to deal with and have been completely omitted of the letter,” he mentioned.

You possibly can watch a video of the trade under:

Altman’s feedback are attention-grabbing — although not essentially due to what they reveal about OpenAI’s future plans. As an alternative, they spotlight a big problem within the debate about AI security: the issue of measuring and monitoring progress. Altman might say that OpenAI just isn’t at present coaching GPT-5, however that’s not a very significant assertion.

A number of the confusion could be attributed to what I name the fallacy of model numbers: the concept numbered tech updates replicate particular and linear enhancements in functionality. It’s a false impression that’s been nurtured on this planet of shopper tech for years, the place numbers assigned to new telephones or working techniques aspire to the rigor of model management however are actually simply advertising and marketing instruments. “Effectively after all the iPhone 35 is healthier than the iPhone 34,” goes the logic of this method. “The quantity is greater ipso facto the cellphone is healthier.”

Due to the overlap between the worlds of shopper tech and synthetic intelligence, this identical logic is now typically utilized to techniques like OpenAI’s language fashions. That is true not solely of the kind of hucksters who publish hyperbolic 🤯 Twitter threads 🤯 predicting that superintelligent AI can be right here in a matter of years as a result of the numbers preserve getting larger but additionally of extra knowledgeable and complicated commentators. As plenty of claims made about AI superintelligence are primarily unfalsifiable, these people depend on related rhetoric to get their level throughout. They draw obscure graphs with axes labeled “progress” and “time,” plot a line going up and to the correct, and current this uncritically as proof.

This isn’t to dismiss fears about AI security or ignore the truth that these techniques are quickly bettering and never absolutely below our management. However it’s to say that there are good arguments and dangerous arguments, and simply because we’ve given a quantity to one thing — be {that a} new cellphone or the idea of intelligence — doesn’t imply we’ve got the total measure of it.

As an alternative, I believe the main focus in these discussions must be on capabilities: on demonstrations of what these techniques can and may’t do and predictions of how this will likely change over time.

That’s why Altman’s affirmation that OpenAI just isn’t at present growing GPT-5 gained’t be of any comfort to individuals frightened about AI security. The corporate continues to be increasing the potential of GPT-4 (by connecting it to the internet, for instance), and others within the trade are constructing equally bold instruments, letting AI techniques act on behalf of customers. There’s additionally all kinds of labor that’s little doubt being achieved to optimize GPT-4, and OpenAI might launch GPT-4.5 (because it did GPT-3.5) first — one other manner that model numbers can mislead.

Even when the world’s governments have been by some means in a position to implement a ban on new AI developments, it’s clear that society has its fingers full with the techniques at present obtainable. Certain, GPT-5 isn’t coming but, however does it matter when GPT-4 continues to be not absolutely understood?