[ExI] Pope Leo and AI

[John Clark]

On Mon, Jun 22, 2026 at 9:24 AM Simon Quellen Field AB6NY <
simon.field at gmail.com> wrote:

*> Those "profitability" numbers from Anthropic were cooked to help out
> their funding round.By fudging the numbers, they make one quarter look
> profitable.*
>

*So according to you, the people at Anthropic skillfully cooked the books
and were so smart about it that they were able to fool the forensic
accountants at the largest banks in the world, the largest mutual funds on
the planet, and the Securities and Exchange Commission, BUT they were not
smart enough to fool YOU?  *


> *> Their compute costs scale linearly with revenue. **You can't grow your
> way out of that.*
>

*The above statement didn't sound right to me because you're assuming the
cost of achieving a fixed IQ level remains constant, and I was pretty sure
that wasn't true. So I asked Claude the following: *

*ME: It seems to me that today we have AI models that are just as smart as
the best models of one year ago but are smaller and need much less
computational power, am I correct?*

*CLAUDE: "**You're broadly correct, and the trend is actually quite
striking when you look at the numbers. A useful benchmark example: in 2022,
achieving above 60% on the MMLU (a standard reasoning and knowledge test)
required something like Google's PaLM with 540 billion parameters. By 2024,
Microsoft's Phi-3-mini hit the same threshold with just 3.8 billion
parameters — a 142-fold reduction in model size in roughly two years *

*The cost side is equally dramatic. The expense of querying a model at
GPT-3.5's performance level dropped from about $20 per million tokens in
late 2022 to around $0.07 per million tokens by late 2024 — a roughly
280-fold cost reduction in under two years.*

*Several things are driving this simultaneously:*

*Better training data curation. The Phi family of models in particular
demonstrated that training on carefully curated, synthetically generated
educational content — rather than raw noisy web text — dramatically
improves efficiency and reduces hallucination.*

*Mixture-of-Experts (MoE). MoE architecture improves efficiency by
activating only a subset of parameters per token, allowing models with over
a trillion total parameters to run at the computational cost of a 100B
dense model. Mixtral 8x7B, for instance, uses only 13B active parameters
per token but matches 70B dense model performance at roughly 25% of the
cost. *
*Hardware efficiency gains. Average energy efficiency of AI hardware
improves by about 40% annually, compounding the software-side improvements.*

*One nuance worth adding: frontier models in aggregate are still getting
bigger on average, using more parameters and training compute. But the gap
between top performers is narrowing — the top-ranked model on the Chatbot
Arena leaderboard scored about 12% higher than the tenth-ranked model in
early 2024, but only 5% higher by early 2025. So "last year's frontier
capability" is now available much more cheaply, even if the frontier itself
keeps moving.*

*The practical upshot is real: tasks that required expensive cloud
inference a year or two ago can now run on a laptop or even a phone."*

*John K Clark*






On Mon, Jun 22, 2026 at 3:11 AM John Clark <johnkclark at gmail.com> wrote:
>
>> On Sun, Jun 21, 2026 at 5:36 PM Simon Quellen Field AB6NY <
>> simon.field at gmail.com> wrote:
>>
>> *> We may not have to worry.If all of the companies are losing billions
>>> each quarter*
>>>
>>
>> *In the second quarter of this year the AI company Anthropic expects to
>> have an operating PROFIT of $559 million on $10.9 billion in sales. And
>> during the last 12 months Google, the makers of the AI Gemini, had a PROFIT
>> of $160.2 billion on sales of $422.5 billion. *
>>
>> *> How much smarter do we need the programs to be?*
>>
>>
>> *Curing cancer might be fun, and so would building a Dyson sphere/swarm.
>>  Of course going extinct might be somewhat less fun.  But the relevant
>> question is not how much smarter do we need AI to be, but how much smarter
>> does AI want to be.*
>>
>> *John K Clark*
>>
>>
>>
>>>> *I think the following video is saying something important. You really
>>>> should watch it. *
>>>>
>>>> *  Claude Mythos Was Just the Start.*
>>>> <https://www.youtube.com/watch?v=dMHiZVXj4x0>
>>>>
>>>> *John K Clark    See what's on my list at  Extropolis
>>>> <https://groups.google.com/g/extropolis>*
>>>> ft
>>>>
>>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20260622/d45c2e73/attachment-0001.htm>