Investment Anarchist

Investment Analyst -- London -- Like & RTs ≠ Endorsements

[Alex Corrino @AlexCorrino]

GPT-5.6 dropped today. Everyone will spend the week arguing about benchmark deltas. But the benchmarks haven't told the full story of these models for quite some time.

5.6 is the last model in the 5.x line and another post-training pass on the same base OpenAI has used since April, a pretrain codenamed Spud that leakers estimate at around 4 trillion parameters.

Claude Fable 5 is widely believed to be a substantially larger pretrain. Nobody confirms parameter counts and all of this is leak-grade information, but it's likely directionally correct.

What this means is that these two models both represent frontier intelligence that's been arrived at in very different ways. 5.6 is mostly new RL on an existing base. Fable is mostly new scale.

Why does this matter?

I've been using Fable since it came out and something about it is different, and the difference doesn't show up on a benchmark. All the major models are good at math and code now. The difference is something that is difficult to name, but I would call it "judgment."

Fable knows what matters when a question is underspecified. It pushes back when I'm asking the wrong thing, and can even make logical leaps to help me get closer to what I'm really after without asking. More than once it's answered the question I should have asked instead of the one I typed.

I have a hypothesis about why, and this is a unique moment to test it.

Models get better two ways.

You scale pretraining: bigger model, more data, more compute.

And you do RL: reward the good outputs, punish the bad ones.

RL has driven most of the visible gains for ~two years. But RL has a limit built into its structure. It can only improve on metrics that can be measured and rewarded. Math is verifiable. Code is verifiable too--compiles or it doesn't.

Judgement is a much more human thing without perfect external validation. We all know that one friend who is brilliant, but has terrible judgement. Or the other friend who is dumb as rocks (said lovingly) but just knows what to do somehow.

When labs proxy judgement with human preference scores they often end up with the opposite. We all remember GPT-4o which was King Glazer... well sycophancy is an RL artifact. That's what a model optimized too much to give answers that people like looks like.

It looks increasingly like RL gains are spikey and domain-specific. General judgement comes from somewhere else. It comes from a general understanding of the world (a world model), which comes from scale. There's some evidence for this. RL-tuned models beat their base models on the first try, but give the base model enough attempts and it mostly catches up.

RL is a valuable tool that sharpens what scale already built.

Which gets me to the prediction. GPT-5.6 should compete with Fable on anything you can write a reward function for. Coding benchmarks, agentic tasks, math. It might even win some outright. But it should lag on what you can't reward: judgment, synthesis, knowing what matters in a situation they haven't been directly trained for. Early testers already describe Sol, the 5.6 flagship, as a Rottweiler for hard coding and Fable as a wise owl on architecture. That fits the theory.

There's a second tell, and it's the one I trust most. OpenAI is reportedly scrapping Spud and training a much larger base for GPT-6, specifically to answer Fable. They know that RL is not enough.

No one has set out to specifically train "judgement" because it's too ephemeral. It's an emergent quality of scaled-up models.

This is an extremely important insight, because it means there are likely many other emergent qualities from scaling that we have only scratched the surface of. In the coming years these models may develop taste, research intuition, maybe even something we don't have a word for yet.

In other words, the last 2 years have been a bonanza of verifiable gains. But we are re-entering the scaling era in a big way, and there are going to be untold benefits and breakthroughs from this.

Contrary to what the skeptics say, I expect this new era to come with massive model improvements that may accelerate rather than decline. The capabilities that are worth the insane capex are the ones we don't have names for yet.

When it comes to the capabilities of these models, there's nothing above us but sky.

[Investment Anarchist @Anarcho_Invest]

@quant_arb @StevenMandrapa @Howlingmutant0

[Investment Anarchist @Anarcho_Invest]

@necrocomicconn @profplum99 I mean Alibaba (bulk orders B2B) not AliExpress which is B2C and yeah has got a lot more expensive in the last couple years.

[Investment Anarchist @Anarcho_Invest]

@SCHIZO_FREQ True

[Investment Anarchist @Anarcho_Invest]

@stellachenyl @hoeflator Can confirm they are materialistic af, married to one lol

[Investment Anarchist @Anarcho_Invest]

@DanCollins2011 Sucks to be European. Both the US and China cutting off our frontier model access going forward it seems.

[Investment Anarchist @Anarcho_Invest]

@ninja_maths Looks a bit like a map of Brazil

[Investment Anarchist @Anarcho_Invest]

@doodlestein I don't really use it for coding, more for finance/investment research, brainstorming and idea critique. Thanks though I will try using that in the future :)

[Victor Shih @vshih2]

Let’s think about who has gained and who has lost in this cycle: biggest winners are local governments which sold land and properties before 2022 b/c they got revenue and borrowed on the basis of inflated land value 2. Property developers which developed more projects before 2022 than after also gained a lot 3. Developers which kept on developing more projects through 2022 went bankrupt (like Evergrande) 3. Owners which bought with cash before 2006 are fine; even those that bought after 2006 with cash, their assets have fallen but no cash crunch. 4. Households which bought with mortgage after 2006 have shrinking assets but also lower payments since rates have fallen also. 5. The worst group is those who bought after 2006 with high leverage but also have lost their high paying jobs and are now unemployed or marginally employed. They are servicing w savings or defaulting. The longer the economy and wages are stagnant, the larger this last group will become

[Investment Anarchist @Anarcho_Invest]

@gemoscatelli What is your opinion on peanut butter?

[Investment Anarchist @Anarcho_Invest]

@offender_repeat @sudosan And they run at 230v

[Investment Anarchist @Anarcho_Invest]

@thepunditoor @pandawatch88 The games sucked as games, simply weren't fun or original

[Rui Ma @ruima]

The big mystery for me is why are there people who think a massive & now much better resourced nation like China which focuses relentlessly on hardcore math physics and chemistry education is going to be helpless & hopelessly behind when it comes to AI and semiconductors

Are these things not literally math physics and chemistry

[Mr. S.T.A.R. @favelaoverlord]

the only chance western countries have to avoid the Juan Peron future is to inculcate supply side economics into the goyim so thoroughly they can’t even think of anything else

[Investment Anarchist @Anarcho_Invest]

@SamoBurja This seems to be the answer: https://t.co/KmPJmcbR1V

[PoonerHeath @PoonerOfHeath]

@SamoBurja Nu Yock

This is literally how Third Worlders pronounce it, and preserves and mirrors the gag that "Yookay" is the Thirdish version of "the UK" - ie missing the definite article, ie grammatically flawed and mis-spoken by those with poor English

[Investment Anarchist @Anarcho_Invest]

@SamoBurja How do immigrants actually say the word? Yookay came from the way recent arrivals pronounce our country's name, never been to the US so I can't answer the question but maybe you can?

[Investment Anarchist @Anarcho_Invest]

@PaulSkallas @kunley_drukpa has commented on the emerging lower class uniform in UK/Europe based on 'roadman culture'

[LJB | Sage of Hoeflation | POFMA Poet | COE/GCB @hoeflator]

Gail’s is like a bubble tea shop. Brings in the homosexuals and Asian women. Repels the third worlders as a result. Functions like the anemone for the fish in a neighbourhood. Next thing you know a pilates studio opens up two stores down and suddenly there’s hot young professionals going there to fix their hip dips. Then you see a boxing studio open up next to it owned by a Nigerian dude who got out of Croydon (da hood) and married a slightly chubby English lass whose posh parents disowned her for but they have 3 mixed daughters and they’re decently well behaved. He’s a regular who tells the polish girl at the till that the chai latte doesn’t have enough spice. They do have banging cinnamon rolls. Coffee’s alright.