Artificial Intelligence thread

[Bellum_Romanum]

So, Anthropic launched Mythos / Fable 5 today. The benchmarks put it at ~65 in Artificial Intelligence, which is about a +5 lead on the previous state of the art (Opus 4.8) and nearest competitor (GPT 5.5), but what I found more interesting was this:

View attachment 176424

Looks like Anthropic is actively trying to stifle competitors from using their models to accelerate their own model development by making the model quietly and intentionally direct you in the wrong direction. This is an interesting (and rather ruthless) move against competing labs and further secures their reputation as a closed lab with no intention of sharing with the world, and every intention of monopolizing intelligence capabilities for their own purposes. As always, Western labs can't seem to avoid mask off moments.

It is also a reminder of the importance of AI sovereignty and independence, as the alternative of relying on the US - as the likes of Europe, Japan, South Korea, etc. are doing - means giving them the right to keep you permanently behind via quietly sabotaging the capabilities of models available to you. The sooner the world is rid of the West's dominance of frontier coding AI, the better.

I don't know why this recent development is even a surprise to begin with. The idea that AI would democratize advance knowledge; establishing a more "human-centered development" which would then firmly cement western style democracy in the world over was a pipe dream for the idiots!

Western AI and anything it introduces to the world will always placed it above the rest while the rest pays rent and sustenance extracting the world from its tangible natural resources that would essentially keep it from developing and competing with the west. It'll will always be a "class based, tiered system."

Sadly for the Anthropics of the world but luckily for the rest of mankind, China and Chinese tech is in their way.

 

[HighGround]

Tested Fable 5 (Mythos with guardrails for public) extensively... not impressed.... while it is better than Opus 4.8 and for sure probably the top model right now, the actual capabilities in real world use are still very very far from any sort of "AGI" and/or "ASI" lets put it that way. and I tested both the web and claude code versions and on MAX intelligent with the 1 million context etc etc. It still couldn't do some simple bug fixes and I still had to hand hold, basically its no where near the level of intelligence that these hyperscalers would need it to be at in order to justify their current IPO valuations... plus on top of capabilities lacking for all the hype it was hyped up to be, every other innocent thing was blocked due to an overaggressive filter. In one instance I started a new chat and pointed it to the anthrophic's own public news announcement and gave it anthrophic's own url link and immediately the chat got paused/halted etc... and even in claude code, things like academic "Pi calculator" stuff that worked under Opus, when continued under Fable, got false flagged as malicious etc etc... basically they added the gating that in its current form makes it less than useless.

I knew this was way overhyped. Mythos is a noticeably more capable model, but "Fable 5" gives it away. It's just Opus 5, but renamed so they can market it better.

 

[9dashline]

I knew this was way overhyped. Mythos is a noticeably more capable model, but "Fable 5" gives it away. It's just Opus 5, but renamed so they can market it better.

There is the full Mythos model that Anthropic keeps for themselves for internal use and shares with CIA/NSA etc,
then there is the Mythos 5 that specifically has cybersecurity, biology, chemistry capabilities capped...

Then there is the Fable 5 that has cybersecurity, biology, chemistry capabilities REMOVED and on top of it they have it rigged up so that it automatically reverts back to Opus 4.8 at the slightest trigger, often false positives

And on top of all of that, the Fable 5 model secretly poisons code bases or use cases that attempt to use it for LLM development, (which is NOT the same as trying to distill it, which they have also already blocked)...

The version they allow the public to use is neutered to the point of being counterproductive and on top of all of this, after June 22nd it will be removed from subscriptions and users will be forced to pay extra money in addition to their subscriptions (even if they were paying the top tier of $200/month) for the "privilege" of using this neutered gated model at $50 per million tokens, it won't distill, it won't roleplay, it won't help you with LLM or AI development etc etc... and it snitches on you to governments/agencies with a mandatory 30 day retention of all chats/convo/data... and its filters are so sensitive you can't even use it for most IT work anymore... forget about AI stuff.

 

[tphuang]

https://twitter.com/i/web/status/2064765491282157970

ByteDance and AI drug discovery getting its own unit.


As for Fable/Mythos 5, it was clearly just over hyped thing that Anthropic is throwing out ahead of its iPO to get more funding. As many people have said, it's just unusable for practical applications.

It is also unwise to underestimate the Mythos that US intelligence agency is probably using to conduct cyber warfare against China.

 

[tphuang]

https://twitter.com/i/web/status/2064790706116522061

my latest on Huawei Atlas-950 Super Node sales & DeepSeek effort to build its own data center, probably using Atlas nodes.