Joseph Peters

AI Forward, Leader, Explorer. President + Writings Thoughts are my own.

[Joseph Peters @jcpcdn]

Here is all you need to know and consider about the US Government shutting down Anthropic's Fable model. Most of it is bad.

TLDR: This is the Archduke Franz Ferdinand moment of politics entering AI.

The US Government effectively shut down Anthropic's latest model, Fable, at 5pm on a Friday. We've entered the government control phase of AI. This was inevitable as the models move from more powerful to extremely powerful. The Fable model was a giant leap forward. Our Data Science and Development team at TQStarling was able to compress 90 days of development into an evening with Fable. This is an example of the exponential leaps we've been talking about with AI.

Ultimately this looks like an overreaction, but there is bad blood between the current US Administration and Anthropic.

There's a backstory here of Anthropic leadership not kissing the ring of Pete Hegseth (Secretary of War). Anthropic didn't want its models used for public surveillance. Hegseth called CEO Dario Amodei names and then designated Anthropic a national threat and put the company on an export control list. No American company has ever received that designation. Usually it's reserved for countries like China that run spying operations directed at the US.
Now fast forward to Mythos, which is the unlocked version of Fable. It can hack companies easily and possibly help create bioweapons. Anthropic is very responsible with this Mythos model. It creates a project called Glasswing to let governments, businesses, and experts test and evaluate it.

Fast forward a few months and they release Fable. It has some pretty powerful guardrails. On any attempt to use it for cyberattacks or bioweapons (yes, this is a real threat), the model stops the user. They have also introduced a 30-day hold on all user history. Ostensibly this is to audit for criminal use and to check whether any safeguards were bypassed.

Another interesting design choice: if someone tries to use Fable (5.0) and the system flags the use, it degrades to Opus 4.8 instead of just blocking. The Opus model is not seen as the same threat, or as capable as Fable, for criminal purposes.

The US government's requirement of Anthropic was to ensure that non-US citizens did not have access to the Fable model. This isn't about geofencing, as there are many non-US citizens within the United States. Anthropic's ability to comply would require not only citizenship checks but also ensuring the model stays in the hands of citizens. This is very hard, if not impossible.

There are some more insidious and hypothetical angles worth considering.

Where did this pushback come from? Could OpenAI or other labs with good relationships with the current US Administration have pushed the government to slow Anthropic down? The power of the model was only apparent to a few, for a very short time. OpenAI has also made some very significant donations to the Trump administration.

Could industry also have pushed for this? This type of capability increase is not good if you're in the SaaS business. This has SaaSpocalypse 2 written all over it.

Anthropic has stated publicly that these were known issues and not severe. If we take them at their word, then this is either the US government overreacting or punishing Anthropic. Or there are some other strings being pulled by competitors or those negatively impacted by the model.

So this is where we are. A turning point in our AI era, with a shift in capability that has national interests at stake.

This also does not bode well for the open-source models of the future. If they follow in the months to come, then we have a much more complex situation. This is also a problem if the Chinese models are only a couple of months behind.

We're in uncharted waters, and it's looking like gale-force winds are heading our way.

If you consider what this means for countries outside the US and China, it is also quite challenging. It creates a have-and-have-not environment. If the benefits compound, then this has real implications for competitiveness. A Mistral in France or a Cohere in Canada are behind in the frontier model race and in the infrastructure to support it (electricity, chips, and data centres). Sovereign infrastructure must be a priority.

This is a story that will continue to unfold, most likely in familiar and unfamiliar ways, when dollars and power are at stake.

[Joseph Peters @jcpcdn]

@itsolelehmann I’m from Canada. We’re in the same boat. Great post.

[Joseph Peters @jcpcdn]

I think it probably because it’s the safest non offensive option. No one really gets upset about this answer. The AI is zen seeking enlightenment. Now if it had said Catholicism or Muslim or Protestant you create a surface area for attack/rejection. Wonder if there are regional differences to the response.

[Joseph Peters @jcpcdn]

@VoiceOfFranky Frank - super interesting. The big question is how do you convince those consuming disinformation that it is disinformation?

[Joseph Peters @jcpcdn]

@paulroetzer Connecting the dots on this seems so obviously bad.

[Joseph Peters @jcpcdn]

@paulroetzer Facts are irreverent when it comes to beliefs and ideology.

[Joseph Peters @jcpcdn]

@Darbybailey @icreatelife This is excellent!

[Joseph Peters @jcpcdn]

@VoiceOfFranky Hi Frank - this and the next slide really paints a desperate picture of the <35 perspective. Take this and Ai Job displacement projections and there is generationally a real problem.

[Joseph Peters @jcpcdn]

@Rainmaker1973 Big Lebowski

[Joseph Peters @jcpcdn]

@adonis_singh Recreated w 4.5. I just added “longer and more dystopian.” Incredible.

[Joseph Peters @jcpcdn]

A lot of progress here.

[John Rush @johnrush]

After seeing these humanoid robot demos, I bet you'll be convinced that all manual labor will be gone to robots.

(even the world's oldest profession will be taken by them).

All 26 humanoid robot demos:

[Alessandro Palombo @thealepalombo]

Everyone thinks Europe is finished. In 20 years, they'll wish they had bought in early.

AGI takes over. Productivity is infinite. Everything is automated.

So, tell me: what actually becomes scarce?

NOT another piece of software — but authentic human experience.

When machines handle everything, what do people crave? Beauty. Meaning. Significance.

And Europe has been accumulating that for centuries.

It's sitting on the most undervalued asset of the AI age:

- 500+ UNESCO World Heritage sites (the US? 25)
- The world's greatest museums
- 50M+ cultural tourists in France alone
- Centuries-old universities, libraries, cafés
- The birthplace of opera, ballet, fine wine

The real arbitrage? Owning land in places machines can't replicate.

In the AI age, people will split into two groups:

- New "landlords" stacking assets
- New "renters" living off AI welfare (UBI, digital credits, whatever comes next)

So where will the new “landlords” want to live?

Not in a sterile AI-optimized city. Not in a hyper-efficient pod.

They'll want Paris, Rome, Vienna. Cities that weren't built for algorithms, but for the soul.

Europe today is like Bitcoin at $10—misunderstood, underpriced, and wildly asymmetric.

I talk to global investors every day. The smartest ones see it:

In an AI-dominated world, the ultimate luxury won't be another software tool. It'll be the ability to feel human.

Europe has that.

What's your take on this?

[Joseph Peters @jcpcdn]

There is going to be some amazing AI art in the future.

[AshutoshShrivastava @ai_for_success]

AI is insane! This is one of the most mind-blowing things I’ve seen in a while 🤯. You just have to watch it.

[Sam Altman @sama]

we put out an update to chatgpt (4o).

it is pretty good.

it is soon going to get much better, team is cooking.

[Joseph Peters @jcpcdn]

Worth a look. Easier times for prompting.

[Alvaro Cintas @dr_cintas]

OpenAI has released a new prompting guide for their reasoning models.

It emphasizes simplicity, avoiding chain-of-thought prompts, the use of delimiters, and when to use them.

Here’s a breakdown and an optimized prompt to have it write like you:

[Andrej Karpathy @karpathy]

I was given early access to Grok 3 earlier today, making me I think one of the first few who could run a quick vibe check.

Thinking
✅ First, Grok 3 clearly has an around state of the art thinking model ("Think" button) and did great out of the box on my Settler's of Catan question:

"Create a board game webpage showing a hex grid, just like in the game Settlers of Catan. Each hex grid is numbered from 1..N, where N is the total number of hex tiles. Make it generic, so one can change the number of "rings" using a slider. For example in Catan the radius is 3 hexes. Single html page please."

Few models get this right reliably. The top OpenAI thinking models (e.g. o1-pro, at $200/month) get it too, but all of DeepSeek-R1, Gemini 2.0 Flash Thinking, and Claude do not.

❌ It did not solve my "Emoji mystery" question where I give a smiling face with an attached message hidden inside Unicode variation selectors, even when I give a strong hint on how to decode it in the form of Rust code. The most progress I've seen is from DeepSeek-R1 which once partially decoded the message.

❓ It solved a few tic tac toe boards I gave it with a pretty nice/clean chain of thought (many SOTA models often fail these!). So I upped the difficulty and asked it to generate 3 "tricky" tic tac toe boards, which it failed on (generating nonsense boards / text), but then so did o1 pro.

✅ I uploaded GPT-2 paper. I asked a bunch of simple lookup questions, all worked great. Then asked to estimate the number of training flops it took to train GPT-2, with no searching. This is tricky because the number of tokens is not spelled out so it has to be partially estimated and partially calculated, stressing all of lookup, knowledge, and math. One example is 40GB of text ~= 40B characters ~= 40B bytes (assume ASCII) ~= 10B tokens (assume ~4 bytes/tok), at ~10 epochs ~= 100B token training run, at 1.5B params and with 2+4=6 flops/param/token, this is 100e9 X 1.5e9 X 6 ~= 1e21 FLOPs. Both Grok 3 and 4o fail this task, but Grok 3 with Thinking solves it great, while o1 pro (GPT thinking model) fails.

I like that the model *will* attempt to solve the Riemann hypothesis when asked to, similar to DeepSeek-R1 but unlike many other models that give up instantly (o1-pro, Claude, Gemini 2.0 Flash Thinking) and simply say that it is a great unsolved problem. I had to stop it eventually because I felt a bit bad for it, but it showed courage and who knows, maybe one day...

The impression overall I got here is that this is somewhere around o1-pro capability, and ahead of DeepSeek-R1, though of course we need actual, real evaluations to look at.

DeepSearch
Very neat offering that seems to combine something along the lines of what OpenAI / Perplexity call "Deep Research", together with thinking. Except instead of "Deep Research" it is "Deep Search" (sigh). Can produce high quality responses to various researchy / lookupy questions you could imagine have answers in article on the internet, e.g. a few I tried, which I stole from my recent search history on Perplexity, along with how it went:

- ✅ "What's up with the upcoming Apple Launch? Any rumors?"
- ✅ "Why is Palantir stock surging recently?"
- ✅ "White Lotus 3 where was it filmed and is it the same team as Seasons 1 and 2?"
- ✅ "What toothpaste does Bryan Johnson use?"
- ❌ "Singles Inferno Season 4 cast where are they now?"
- ❌ "What speech to text program has Simon Willison mentioned he's using?"

❌ I did find some sharp edges here. E.g. the model doesn't seem to like to reference X as a source by default, though you can explicitly ask it to. A few times I caught it hallucinating URLs that don't exist. A few times it said factual things that I think are incorrect and it didn't provide a citation for it (it probably doesn't exist). E.g. it told me that "Kim Jeong-su is still dating Kim Min-seol" of Singles Inferno Season 4, which surely is totally off, right? And when I asked it to create a report on the major LLM labs and their amount of total funding and estimate of employee count, it listed 12 major labs but not itself (xAI).

The impression I get of DeepSearch is that it's approximately around Perplexity DeepResearch offering (which is great!), but not yet at the level of OpenAI's recently released "Deep Research", which still feels more thorough and reliable (though still nowhere perfect, e.g. it, too, quite incorrectly excludes xAI as a "major LLM labs" when I tried with it...).

Random LLM "gotcha"s

I tried a few more fun / random LLM gotcha queries I like to try now and then. Gotchas are queries that specifically on the easy side for humans but on the hard side for LLMs, so I was curious which of them Grok 3 makes progress on.

✅ Grok 3 knows there are 3 "r" in "strawberry", but then it also told me there are only 3 "L" in LOLLAPALOOZA. Turning on Thinking solves this.
✅ Grok 3 told me 9.11 > 9.9. (common with other LLMs too), but again, turning on Thinking solves it.
✅ Few simple puzzles worked ok even without thinking, e.g. *"Sally (a girl) has 3 brothers. Each brother has 2 sisters. How many sisters does Sally have?"*. E.g. GPT4o says 2 (incorrectly).
❌ Sadly the model's sense of humor does not appear to be obviously improved. This is a common LLM issue with humor capability and general mode collapse, famously, e.g. 90% of 1,008 outputs asking ChatGPT for joke were repetitions of the same 25 jokes​. Even when prompted in more detail away from simple pun territory (e.g. give me a standup), I'm not sure that it is state of the art humor. Example generated joke: "*Why did the chicken join a band? Because it had the drumsticks and wanted to be a cluck-star!*". In quick testing, thinking did not help, possibly it made it a bit worse.
❌ Model still appears to be just a bit too overly sensitive to "complex ethical issues", e.g. generated a 1 page essay basically refusing to answer whether it might be ethically justifiable to misgender someone if it meant saving 1 million people from dying.
❌ Simon Willison's "*Generate an SVG of a pelican riding a bicycle*". It stresses the LLMs ability to lay out many elements on a 2D grid, which is very difficult because the LLMs can't "see" like people do, so it's arranging things in the dark, in text. Marking as fail because these pelicans are qutie good but, but still a bit broken (see image and comparisons). Claude's are best, but imo I suspect they specifically targeted SVG capability during training.

Summary. As far as a quick vibe check over ~2 hours this morning, Grok 3 + Thinking feels somewhere around the state of the art territory of OpenAI's strongest models (o1-pro, $200/month), and slightly better than DeepSeek-R1 and Gemini 2.0 Flash Thinking. Which is quite incredible considering that the team started from scratch ~1 year ago, this timescale to state of the art territory is unprecedented. Do also keep in mind the caveats - the models are stochastic and may give slightly different answers each time, and it is very early, so we'll have to wait for a lot more evaluations over a period of the next few days/weeks. The early LM arena results look quite encouraging indeed. For now, big congrats to the xAI team, they clearly have huge velocity and momentum and I am excited to add Grok 3 to my "LLM council" and hear what it thinks going forward.

[Joseph Peters @jcpcdn]

@maxfawcett Ontario accounted for 38% of national GDP in 2023. Quebec was next at 19.8% Sounds like you have some skewed perspective of wealth.

[Joseph Peters @jcpcdn]

Hey @fatherjohnmisty just to let you know the presale queue is faulty. Waited to get to my turn. When I got to my turn it put me back to the bottom of the list with 500 new people in front of me. Sad that this keeps on happening to fans.

[Joseph Peters @jcpcdn]

@milkkarten @washingtonpost @drewharwell @KamalaHQ This is really excellent.

[Joseph Peters @jcpcdn]

@TemplesOfficial Amazing. Super cool story.