AI #176 Part 1: Doing It Live

Zvi Mowshowitz

Jul 09, 2026

Enough things added up that this week is getting split into two parts.

Then on Monday, if all goes as I expect, we’ll cover OpenAI’s Sol, aka GPT-5.6.

OpenAI also gave us an upgraded voice mode, which I haven’t tried out but early reports are that it is a step change.

AI writing, especially Claude writing, is becoming more prominent and harder not to notice, and increasingly a tough read when encountered in the wild. Does anyone care? Or are those who care the weird ones here?

This week saw an excellent paper, which I cover in No Space Like J-Space.

Technically we also got Grok 4.5.

Table of Contents

Language Models Offer Mundane Utility. A whole new world.

Language Models Gain Unexpected Affordances. Wait, you can just do that?

Language Models Don’t Offer Mundane Utility. Things get old.

Pay The Man His Money. You have a few more days with marginally free Fable.

Huh, Upgrades. Anthropic raises API platform limits.

Grok 4.5 Exists. It might be okay for its price.

F*** It We’re Doing It Live. OpenAI gives us a big upgrade to voice mode.

On Your Marks. Games are the ultimate benchmarks.

Better Call Sol. Coming soon! Get hyped.

Get My Agent On The Line. Fable makes choices, Replit continuously learns.

Deepfaketown and Botpocalypse Soon. Stop it with the AI-written drivel, please.

Fool Me Twice. I won’t get fooled again unless you put in a little effort.

I Like Your Style. Alas, I might be the weird one. Perhaps no one else cares.

Enough With That Style. You’re absolutely right — this is getting old fast.

Fun With Media Generation. F1 as well-executed, zero-perplexity non-AI slop.

Copyright Confrontation. Hugging Face not beating the rumors.

Cyber Lack of Security. Pliny goes on the offensive, I mean only for white hats.

A Young Lady’s Illustrated Primer. If given the chance, the entire class cheats.

They Took Our Jobs. Those who use AI create and also destroy jobs.

Get Involved. AI protest in SF, microgrants, Palisade Research.

In Other AI News. Never stop being a jackass for safety.

Show Me the Money. Coefficient Giving grants $160 million to Resolution.

Bubble, Bubble, Toil and Trouble. AI as an ordinary systemic financial risk.

Language Models Offer Mundane Utility

Ethan Mollick: I had Fable build another thing I always wanted, a full procedural fantasy kingdom generator with economics, trade routes, population growth, wars, lineages, and occasional dragons. First, I worked with it on a plan, then it made it.

You can play it here.

Also signs and portents, royal processions, mule trains, bandit camps, tiny sheep, rivers, plagues, assassinations, marriages, fields, natural resources, and other stuff.

I heard it didn't work great on phones, I told Fable. Now it does. Mac trackpads fixed. I should just have Fable monitor the thread for bug reports and solve them.

Use an AI face tell analyzer for WSOP coverage on ESPN. Presumably the next step is that poker players train against the tracker.

Fable is my new trusted fact checker and copy editor. One could have previously used Opus 4.8 or GPT-5.5, and probably I should have, but they didn’t cross the threshold where I felt they justified the activation energy. Fable absolutely does and I assume Sol (or Sol Pro) will as well. It is likely one should now use both.

The marginal value of output you get from a superior LLM can be worth quite a lot. In the example here, about $165k was spent on Claude for a porting job that would have taken three top level years of work. Yes, you could try and do it cheaper, and if possible you should do that, but if you can offer a better product you can rake it in.

The danger with such calculations is confusing costs and benefits. The cost of doing it by hand does not tell you whether the result is valuable. In this case, it is clear that it was.

Dwarkesh Patel: Seems to suggest that if it stops being the case that there's 3 labs which are all roughly equally good, competing each others margins away, the provider of the best model could probably get away with charging *a lot* more than they currently are.

We are now down to two labs offering top models, and those two models are distinct from one another. So pricing power is going up for now, not down.

Language Models Gain Unexpected Affordances

A fun theme is ‘Fable uses affordances the user did not realize it had.’

So far all of the examples I have seen in the wild have been harmless in practice, but there’s very much a ‘wait no I didn’t tell you to do what now?’ and a ‘wait you can just do that?’ that is growing increasingly unsettling. Expect its surface area to expand with time, and for the things AIs figure out how to do to grow increasingly surprising.

0.005 Seconds (3/694): my wife asked me to clone a site for her for work and in the process of doing so claude appears to have logged into their unsecured admin portal to screenshot the layout.

Alex Godofsky: I asked Fable to write a discord scraper for a small task, and when I told it "okay let's fill in my auth token" it said "sure thing boss I'll go extract it from your browser cookies" and I was all "wait wait stop what I didn't mean that".

Vivienne Bellerose: This sort of thing happens CONSTANTLY

Liora has Fable proactively monitoring her downloads folder, and she wonders about it in the future using the camera.

Here’s a more fun new affordance from a different project.

Amir Zamir: Turns out it's possible to generate videos that maximally excite an arbitrary brain region using a simple search-based algorithm. It's a fully computational approach, so it's another way to speculate what a brain region represents, alongside other neuroscientific methods.

Select an arbitrary brain region->algorithmically generate a video that jacks it up. See the visuals on the webpage https://nevo-project.epfl.ch. In silico (for now).

Yingtian Tang: Website: https://nevo-project.epfl.ch arXiv: https://arxiv.org/abs/2607.02317 Model card: https://huggingface.co/epfl-neuroai/NEvo

Yingtian Tang: How it works:

Given a target ROI, we evolve text prompts over a structured search space (30 attribute categories, 614 options). The optimization loop:

prompts → videos videos → predicted ROI response ROI response → evolved prompts

One should think seriously about the implications of this, and what a sufficiently advanced AI could do to a human brain using advanced versions of this technique.

Language Models Don’t Offer Mundane Utility

Raymond is impressed by Fable’s first story, then notices it writes similar stories over and over again. Yeah, the models be like that, especially if you don’t switch up context. Also most human authors be like that.

Whereas Eliezer Yudkowsky is not impressed in absolute terms on fiction and plot writing, seeing giant mistakes, although it is still a big step up from old models. He does find it a large step up in decision theory intelligence.

Sam Morril not only doesn’t use AI to help write jokes he mostly, like many comedians, doesn’t use any screens at all, to get rid of all distractions.

Pay The Man His Money

Claude: We're extending access to Claude Fable 5 on all paid plans through July 12.

Eliezer Yudkowsky: I might otherwise resent this but I think the apparent deadline in fact caused me to get around to doing various stuff, and therefore I have no right to complain.

j⧉nus: why would you have stopped using fable if they werent on the subscription? api costs too high or you cant be bothered to use anything but the app or..?

Eliezer Yudkowsky: API costs large enough that I notice, more like 100x subscription than 10x.

Huh, Upgrades

Anthropic raises Platform API limits and simplifies its tiers.

Grok 4.5 Exists

It has 1.5 trillion parameters. Price is $2/$6, or $4/$18 for the fast version.

It was trained in large part by Cursor, so it is kind of a hard reset.

It claims some good benchmarks. As in, there are four good benchmarks.

They shared almost nothing else.

In case there was any doubt, yes, Pliny jailbroke it.

Those scores mean Grok 4.5 is almost certainly a large improvement in coding over previous Grok models, but choosing to present it in this way suggests it will rather soundly underperform what these benchmarks suggest. If they had a model on the level of Opus 4.8 and GPT-5.5, they’d be louder about it. The lack of outside reactions reinforces this.

It certainly is not going to be competitive with GPT-5.6-Sol or Fable. The good news for SpaceX is that this is cheaper, so it might have its uses. But given the track record, I’m going to wait for positive signs before I do anything about it.

F*** It We’re Doing It Live

OpenAI introduces GPT-Live, which they call a new generation of voice models for natural human-AI interaction, including a sense of time and transition. If good enough, this can plausibly be a step change, where suddenly it is good enough to talk to.

This official thread has some videos of people talking to it.

Some people can’t wait for this to be good enough to shift their baseline mode to voice. I am very much not that, I believe text is typically superior to even ideal voice.

My brain cannot comprehend wanting to code via voice, yet many swear by it.

Either way, certainly voice has its niches. Sometimes it is annoying to type.

Sam Altman (CEO OpenAI): GPT-live (next-generation voice) launches today in ChatGPT.

it feels magical and 'real'.

i have always preferred typing to talking to an AI, now i think that's going to shift.

Riley Coyote: gpt-live voice is a very solid improvement

On Your Marks

EldenRingCorruptedSaveFileBench, Fable scores 100% up from everyone’s 0%.

July Fable underperforms June Fable on many benchmarks, reflecting that it more often falls back to Opus 4.8. APEX-SWE is one example, where roughly half its advantage over Opus 4.8 was lost.

Epoch AI introduces EBR-bench, where AIs play a board game Earthborne Rangers and try to learn from their mistakes via a notepad. None of the AIs improve over time, and even a full strategy guide only modestly helps. The models mostly don’t explore. The game looks cool but is out of print and I didn’t see an online version. Models struggle with deckbuilding and also tactics.

Better Call Sol

GPT-5.6-Sol will be available later today, along with Terra and Luna.

Until then, here is some early hype.

If the hype is real, it would be a hell of a trip. When not tripping the classifiers, Fable is clearly far superior to every previously existing LLM across the board. If Sol is indeed often even better than that? Yowsers.

But as Roon points out, those with early access are a highly biased group. Give it time.

tylercowen: GPT 5.6 has *excellent* judgment, as an early tester I will vouch for this.

Ethan Mollick: I was an early tester of GPT-5.6 Sol. I was asked to not share demos until after launch but it is a very good model.

It is of similar ability, but quite different feel, than Fable. Fable wants to go off and do work on its own pace, Sol is faster but works with you in steps more.

I found myself switching between Fable and Sol depending on task. Sol for back-and-forth tasks, especially when I had not yet figured out what I needed exactly, Fable for very long tasks where I could define what I wanted, and Sol Pro for really hard problems.

Fable feels very different than Opus. GPT-5.6 feels like a part of the GPT-5 family. I developed a very complex set of heuristics about when to use which. Fable was often “smarter” but was also too self-directed for some work, while that characteristic was perfect for others.

Ethan Mollick: My big takeaway is that both Sol & Fable represent jumps over previous models and have opened a large gap with the next-best AIs. People will have preferences for one or the other, but if you doing any work where better intelligence matters, those two models are your only choices

Dan Shipper: GPT-5.6 is a much better writer than Fable.

It c

[truncated for AI cost control]