Sam Altman’s Attack, Amazon vs. Starlink, and What Opus 4.7 Actually Means | #248

Peter Diamandis (0:00)

A 20-year-old Texan threw a Molotov cocktail at Sam Altman's San Francisco house. Suspect was on something called the official Paws AI Discord server list. The state of Maine passed the first ever statewide data center ban in the United States. Social unrest coming as a result of people's fear and people not getting jobs. Only 23% of the public is optimistic about AI.

Dave Blundin (0:24)

99% of the people you bump into on the street are underreacting and unaware.

Salim Ismail (0:29)

If you don't want to use it, fine, let other people use it and get the benefits of it.

Peter Diamandis (0:34)

Anthropics Opus 4.7 dropped.

Alexander Wissner-Gross (0:36)

It is moderately interesting. Is it mythically interesting? No.

The new guidance is use prompts. Use prompts for everything. The problem is it's sort of an outburning effect where I want mythos access.

Peter Diamandis (0:52)

Amazon and Apple team up to compete against Starlink.

Alexander Wissner-Gross (0:56)

I would bet that Apple in short turn ends up pitting Amazon, the new global star owner, against SpaceX Starlink.

Peter Diamandis (1:04)

Elon does not stand still.

Dave Blundin (1:08)

Now that's a moonshot, ladies and gentlemen.

Peter Diamandis (1:13)

So as we started recording this episode today, Anthropics Opus 4.7 dropped. So we wanted to do a quick pick up, insert it here at the top of the show to discuss what is Opus 4.7? How does it compare to 4.6?

Mythos, of course, we're here with our resident genius on all benchmarks, Alex Wissner-Gross.

Alexander Wissner-Gross (1:33)

It is moderately interesting. Is it mythically interesting?

Is it incrementally interesting? No, it's a solid release. I've been using it for the past few hours. I ask it to, my standard go-to, as loyal viewers of the pod may recall, is asking it to generate a cyberpunk first-person shooter game design that's visually stunning, and it generated something that was visually stunning. The benchmarks are interesting.

The bio benchmarks in particular are interesting. It's a solid release. It's probably, if I had to guess, a further post-training of some existing model. Could be a distillation of a larger model. Could be a distillation of mythos, potentially. Not quite clear. But I would say it is a solid point release of Opus. And the problem almost is an expectation anchoring one, having seen the eval results for mythos or mythos, as you like to say.

Peter Diamandis (2:29)

I like calling them mythos, yes.

Alexander Wissner-Gross (2:32)

The problem is it's sort of an Osborning effect where I want mythos access. Give me mythos access. And then when you compare the Opus 4.7 benchmarks with mythos, you feel, I don't know, a sense of-

I was going to go with ennui, but you can pick your own superlative here. So if you look at migration, so I think it was particularly instructive to look at migration notes between 4.6 and 4.7.

The biggest change that I could see is that all of the dials and hyperparameters that used to be present in 4.6 and earlier, like temperature, for example, there's no temperature knob anymore. I think that's really instructive. There's no ability for reasoning to control the number explicitly of reasoning tokens that are allowed by 4.7. Now everything is down to a handful of categorical settings where extra high reasoning is the recommended default, maximum mode, and then there are lower reasoning efforts than that. And I think we're seeing, in some sense, an end of an era where the earlier controls that we used to have, remember back in the good old days, like six months ago, it used to be possible to turn the temperature of a frontier model down to zero to get quasi-deterministic behavior, for those who care about that sort of thing. No longer possible. Now you're just told in the documentation, you want determinism, forget about it. Temperature equals zero never was deterministic in the first place. Now everything, the new guidance is use prompts. Use prompts for everything. Prompts are the new dials and the new hyperparameters. And if you want something like, say, a reasoning model to emit guidance regarding its reasoning trace every three seconds, now you're supposed to ask for it in natural language. The knobs are gone.

Peter Diamandis (4:17)

Dave, you're more excited about the model than a lot of people are.

Dave Blundin (4:22)

Yeah, well, it's interesting, you know, it dropped three hours ago, so I've been using it for three hours now. So that's, but right out of the gate, you know, it dropped into cursor just fine, just click and go. It's in Claude Cowork just fine, click and go. But then Claude Code, it said, well, you know, you got to update your terminal, you got to update your node. So, you know, I noticed in computer use, it's notched way up in its score, and it had no trouble manipulating my computer to install itself, installing a new version of node, installing a whole new terminal that I didn't have on the machine before. And I don't think 4.6 would have done that. Also, I kicked off a whole bunch of agents.