📩 ah so, what's consciousness?

P.S. Think we’re going to be moving the newsletter to Monday mornings, gives everyone (mostly me) some time to breathe.

Ello chaps,

Here’s the TLDR:

• Facebook ads (and some social media content) are back

• Claude Cowork now runs even if you shut your laptop

• ChatGPT is back with their 5.6 model (reported to be better than Fable), OpenAI is getting sued by Apple, Grok (and SpaceX) have a fairly good model, but at a much cheaper price.

• And exploring consciousness in a (crazy) world of AI.

💼 1. What’s new in the biz

Facebook ads are back on. Launching new campaigns and changing up the formats — one track for people wanting the one-click AI Brain setup, another for more sophisticated custom workflow builds.

I'm posting more on socials. Been stuck in a bit of a dilemma: I want to put out genuinely good content, but the only way to actually get better at it is to practice in public. Small silver lining, if only a couple hundred people are watching right now, there's not a lot of brand damage to be done while I get the reps in. The goal is to post 4/7 days a week. (Realistically, let's just get through this week first lol) Longer term, I'd like to keep it going for the whole year. It's turned into a genuinely useful, almost cathartic exercise, forcing whatever I've learned recently into a short, explainable format sharpens the thinking.

The newsletter's almost at 500 subscribers, which is awesome, and a few of those came from people sharing it around with no prompting from me (if you’re one of those people, I actually love you).

On the product side: we've been heads-down building the AI agent that lives in your iMessage, with a command centre that sits alongside it. Six users on it right now, and we're planning a serious round of ads over the next month to push it further. This is the real work of productising the AI brain, not just building it for ourselves.

If you know someone that is a time-poor operator (or you are one yourself) that wants to set up their AI brain in 1-click, reach out and I’d love to show you what we’re building.

🤖 2. What’s new in AI

Anthropic and OpenAI are handing out $500k in free credits.

Microsoft is cutting AI costs.

Two contradictory-looking things happened in the same window, and together they tell an interesting story.

Anthropic raised its no-equity credit offer to YC startups from $30k to $500k; OpenAI matched it, plus an optional $1.5M more if you'll give up equity. Some founders are reportedly stacking combined offers past $3 million. That's two labs burning serious money just to keep developers building on their API.

At the same time, Microsoft, the single biggest commercial distributor of both companies' models, is routing routine Copilot tasks in Excel and Outlook away from OpenAI and Anthropic entirely, onto its own in-house models, purely to cut cost. Mustafa Suleyman has said the goal is to eventually get Anthropic spend to zero for anything that doesn't need frontier-grade reasoning.

This is very much what I was talking about last week - where the models are becoming commodities, and more often than not the commodities are becoming to expensive to justify the use of.

OpenAI just shipped GPT-5.6, and it's aimed squarely at Anthropic.

OpenAI launched its new 5.6 models - the competitor to the Fable model.

Three models: Sol (the flagship), Terra (the mid-tier), and Luna (the budget option), all public since July 9 after Washington sat on the release for five weeks over cybersecurity concerns. The pitch is efficiency as much as intelligence. Altman says Sol is 54% more token-efficient on coding tasks, and OpenAI's own benchmarks have it beating Claude's Fable 5 while using half the tokens and costing about a third less. Vendor benchmarks, so add salt (same rule I applied to Grok below).

Two things are worth taking seriously though. Pricing lands at $5 in / $30 out per million tokens for Sol, down to $1 / $6 for Luna, which keeps the price war from the story above rolling. And within a day of launch, a UK government agency reported "universal jailbreaks" that unlock the model's offensive cyber capabilities: the same class of flaw that had regulators leaning on Anthropic earlier this year. The frontier is getting cheaper and more capable at the same time, and the capability cuts both ways.

Cheeky dub for OpenAI who feels like has been lagging a bit in this.

Same week, Musk shipped Grok 4.5. It's SpaceXAI's first release since xAI got folded into SpaceX, the combined company went public, and they bought Cursor (whose coding data went into the training). Musk's calling it "Opus-class", but even SpaceXAI's own benchmarks show it falling just short of the frontier models, so the honest read is: decent, not the best. The price is actually more interesting. $2 in / $6 out per million tokens, against $5/$25 for Opus and $10/$50 for Fable 5. That's a fraction of the price for most of the capability, and for high-volume workflows where good-enough is genuinely good enough, that trade starts to look very sensible. It's the Microsoft story again from the other side: once models are interchangeable for routine work, you buy on price.

Also been using Fable 5 this last week. Anthropic pushed the deadline before it goes back to metered API pricing out to July 12. It's good, but the best use cases are still coming from planning rather than raw execution. Where it's actually been good is fixing up a lot of bugs in the codebase. We’ve been able to one-shot a lot of the bugs that have been coming up in development for

Apple is suing OpenAI for trade secret theft. Filed Friday, and it names two former Apple people including Tang Tan, who left to run OpenAI's hardware. Apple's claim is that OpenAI systematically poached its employees and coaxed confidential designs and components out of them to build its consumer device. OpenAI's response: "We have no interest in other companies' trade secrets." Honestly the whole story is kind of insane and would recommend you having a read here.

Apple sues OpenAI, alleging artificial intelligence company stole trade secrets

I think its unfortunate because these AI companies are already losing trust so quickly amongst the masses. And its not exactly like OpenAI has a clean reputation…

Why is OpenAI planning to become a for-profit business and does it matter?

Note: a quick framing on all the token pricing you'll see in these announcements. A token is the unit AI gets metered in, roughly three-quarters of a word (this sentence is about 20 of them). "Per million tokens" sounds abstract until you convert it: a million tokens is roughly 750,000 words, call it ten novels. The models released this week charge somewhere between $6 and $50 to write that much, so the premium end costs about what one hour of one employee does (if you are of course using a model that much), and the cheap end costs less than a Sydney coffee. Input is what the model reads, output is what it writes back, and output always costs more because that's where the actual generation happens. The way to think about it: tokens are the kilowatt-hours of intelligence.

The interesting part? The bill for cognition keeps falling every quarter.

🔧 3. Builder's hack: run Claude with your laptop open.

Most people run Claude locally and never think past it. The better move is running it in the cloud instead. Claude's released Cowork, which runs from your phone as well as your desktop, so the work keeps going whether you're at your laptop or not.

(Yes that means you don’t have to walk around with your laptop open lol).

The concrete difference: routines can actually run without your computer needing to stay open. Locally, a scheduled task dies the second you close the laptop. In the cloud, it just runs, the morning brief, the end-of-day wrap, whatever you've set up, fires on schedule whether you're at your desk or not.

I think a bigger pattern that I’m noticing is that AI now makes it so much easier for onboarding to be done proactively - the time to value can be reduced so much.

As such, my thesis is that: proactive AI is where this is all heading. The second you sign up for a tool, it should already be taking action to help you, onboarding you itself, adjusting as it learns how you work. Onboarding and teaching people how to use a new tool is still a massive friction point, and it shouldn't be. The tools that win from here are the ones that just get to work the moment you sign up, instead of handing you a blank canvas and a manual.

That’s part of what we’re trying to build.

 🧭 4. Differentiated thoughts: in a crazy world of AI, what is consciousness?

At university I studied a course called Truth, Meaning and Language. Dense stuff. Wittgenstein, Frege, Bertrand Russell. Lot of reading, generally a stiff course that (I think) could have been condensed quite extensively.

In it we looked at trying to understand, what is truth, what is meaning and what is language? And how do each of these help us to understand the other, better.

In it, we looked at Wittgenstein. Early Wittgenstein had this idea called the picture theory of language: words are a conduit for mapping out a representation of the world. Language, on this view, can be broken down analytically. That idea spawned a whole branch of philosophy, analytical philosophy, built on it.

Thirty years later, Wittgenstein reneged on his own theory. His later position: meaning is use. You understand what something means not by picturing it, but by how the word gets used. Take "the ball is red." Red is an arbitrary label for an arbitrary object. Nobody knows a ball is red in any deeper sense, we just collectively agree to call it that. A child points at a red ball and says "orange." He only learns it's actually red when a parent corrects him: no, that's red. Meaning arrives through feedback, not through some private act of perception.

That's the same mechanism large language models learn through. Human feedback, correction, reinforcement, is what took these models from pattern-matching toys to something that argues, plans, and reasons. So it's a bit rich when people wave AI off as a hobby or a toy. It's learning meaning in essentially the same way we do.

Which is what makes Anthropic's recent research interesting. They found Claude has something they call a J-space, a set of internal patterns that function like a global workspace in the human brain. It's reportable (Claude can describe what's in it when you ask what it's thinking), controllable (Claude can deliberately activate it on request), and it causally shapes multi-step reasoning rather than just decorating the output after the fact. It emerged on its own during training.

Crazy part? Nobody designed it in.

The part that is deeply fascinating is the exploration of the subconscious.

Researchers showed Claude a passage written in Spanish, then reached into the J-space and swapped the "Spanish" pattern for "French". Ask Claude what language the passage is in and it says French. Ask for a famous author who wrote in it and it switches from García Márquez to Victor Hugo. But ask it to simply continue the passage, and it writes flawless Spanish, completely unaffected. The label changed but the LLMs ability to execute Spanish did not. Deliberate tasks consult the workspace, while deeply practiced ones run underneath it automatically, the way you can speak grammatically all day without once thinking about grammar. Which is to say: the model has something that behaves like a subconscious.

It then also conducted ethical tests to see if Claude behaved differently for when the J-space existed vs when the J-space did not exist. So they posed an ethical dilemma of Claude acting as an email assistant and finding out about an affair. What was super interesting was that in the J-space they found that Claude could see it was being tested and acted differently. And accordingly, acted more ethically. Whereas, when they removed the J-space Claude just reacted subconsciously, often in some simulations, threatening blackmail against the antagonist in the example (see below).

There is more to unpack and you can read the rest of the writing here (and I would strongly encourage you to do so). Anthropic's writeup is worth reading directly

None of this proves Claude is conscious. Anthropic is careful to say that. What it shows is something narrower and still strange: access consciousness, the capacity to report on and reason with your own thoughts.

Fireship's take on it is worth the seven minutes too.

So where do I land? Not on "Claude is conscious." Nobody can say that, and Anthropic doesn't. But I keep coming back to this: to rule consciousness out, you have to know what consciousness requires, and nobody on Earth knows that. The confident "obviously not" crowd is running on intuition about what silicon can't do, not on an argument.

I think the core premise that people have an issue with is that LLMs are not generalist enough and have no sensory capabilities. But if those are the only two constraints, I think they will generally be solved. LLMs will become good enough, across a number of domains that they will inevitably become more general. And we’re already seeing the rise of humanoid robots to accompany these LLMs.

Ultimately, here’s what we actually have. A system that learns meaning the way Wittgenstein said we do, through use and correction (RLHF) rather than some private inner act. And now a system that grew a workspace for its own thoughts, one it can report on and steer, without anyone designing it in.

I'm not saying the next property to emerge will be experience itself. I'm saying this stuff is getting crazy and it might be time to reconsider our core premises are around what we feel are axioms of the world.

🎶 5. Carve outs

• Sports:

  • Crazy week in AI, but the FIFA World Cup stays undefeated as one of the all time sporting events.

• Music:

  • Recently bought the vinyl covers of the Beatles, the Bee Gees, the Beach Boys and Elvis. Vinyl is cool.

Until next time,

Adios,

Alex