Agostinho Serrano

Prof. @UCS/BR | Shaping the Future of Learning with GenAI | 20+ yrs in EdTech Research & PhD/MSc Mentorship | AI, Cognition & Innovation | Open to Collaboration

[Agostinho Serrano @EducatingwithAI]

Humans will always be necessary.

GenAI doesn’t generalize beyond its training data as LLMs “walk” probabilistic trajectories within a pre-configured possibility space.

They don’t “want” anything not already encoded in patterns. Humans remain the architects: introducing genuine teleological novelty, detecting ontological incompleteness (“something is missing” even when criteria are met), assuming ethical responsibility, and reconfiguring the network itself.

LLMs are potent mediators, mediators without their own project. The paradox: the more powerful they become, the more humans specialize in meta-translation.

We don’t compete in execution.

We sustain the teleological vector. The Greek ἀρχιτέκτων knew the whole while τέκτονες executed parts. Same function, new substrate. The architektōn is back.

[Agostinho Serrano @EducatingwithAI]

@gregmushen @signulll Yeap, this is a major, big time major win for @sama .

They just need to have a 5.6 that is close in benchmarks with Mythos (they can even Benchmaxx it as we speak).

They have mythos-like, RELIABLE model.

Something that isn't turned off when you're building a huge project.

[Agostinho Serrano @EducatingwithAI]

@tszzl Given the current banishment of the models, the ONLY solution is actually a sophon-like training for all sensitive Mythos output (including sending the IP to NSA)

[sdmat @sdmat123]

Anthropic

[Agostinho Serrano @EducatingwithAI]

@apples_jimmy @AnthropicAI Either this is a 6D marketing campaign or he really shoot himself in the foot.

[Agostinho Serrano @EducatingwithAI]

@EricTopol @EvidenceOpen @UpToDate I think this wasn’t anticipated 1-2y ago when Google launched their first model trained with medical data, but quite sincerely, whomever has followed the quite consistent trend knows that general models perform better in any specialised field (see Mythos/Security)

[Agostinho Serrano @EducatingwithAI]

I have been bewildered by Opus 4.6 / Fable 5 a number of times.

Please make no mistake, it’s an alien intelligence. It’s an co-author of ideas, a true cognitive partner. I’m as we speak preparing a paper on the results I shared in your post.

But I still think we will be there outside and above “the loop”

[Agostinho Serrano @EducatingwithAI]

@RifeWithKaiju I still sustain it.

Our greatest advantage as a species is since using Peebles to count cattle is to create tools and leverage them, leaving our brains to complimentary tasks.

We will find what complimentary means with Mythos like and superior models, at least that’s my bet.

[Agostinho Serrano @EducatingwithAI]

@RifeWithKaiju

[Agostinho Serrano @EducatingwithAI]

@RifeWithKaiju

[Agostinho Serrano @EducatingwithAI]

@RifeWithKaiju I've been toying - after noticing how interesting Claude considerations with some academic interactions (4.6 but not 4.7/4.8 and now again with Fable5). I've seen 4 different progressive patterns.

[Agostinho Serrano @EducatingwithAI]

Fable 5 is otherwordly. It reads minds, it can change entirely how you adress a problem. It's a separate entity, not an obedient slave. Incredible insights for Educational Research. I can literally ask which theoretical referential would be best to what "emerges" in the data and it not only suggest, but considers how I can, with my background, use it suggesting papers from people in my field that have "translated" the referential from Educational Philosophy to Science Education (last part unprompted, like it tells you theres a bug in your code while you ask it for optimization).

[Agostinho Serrano @EducatingwithAI]

@lillysharples Tacit knowledge, Yeap. Or like they like to call “ambiguity” (tacit knowledge is what we need to solve many problems that have a lot of ambiguity).

[Agostinho Serrano @EducatingwithAI]

Eles estão super sensíveis a qualquer coisa biológica porque classificaram Mythos/Fable como tendo risco CB-1 mas chegando em CB-2 (o que levaria a não oferecer o modelo ao público sem guardrails pesados); em testes o modelo permite biólogos gerais (não especialistas) a terem a capacidade de especialistas em qualquer campo incluindo bioterrorismo. Como a possibilidade de prompt injection é real ao oferecer ao público amplo, travaram muito o modelo. (System card)

[Agostinho Serrano @EducatingwithAI]

@DaveShapi Just a suggestion: many reports memory is messing the very strick guardrail router.

Perhaps incognito mode works for you.

I had ZERO refusals to anything I threw at him (I don’t work in biotech or security anyway).

[Agostinho Serrano @EducatingwithAI]

@0xchromium That’s exactly how I have been doing.

[Agostinho Serrano @EducatingwithAI]

Frankly, it’s not that hard.

I’m 55, my wife 47.

It’s consistent training (45min, 4x/week) + reasonably good sleep + feeding.

Variety of fresh fruits, vegetables, meat.

The keyword is discipline and preference for healthy food.

I reached 8% bf without much effort.

It’s a lifestyle change, even the free meals consist of healthy food:

Gnocchi made with purple-fleshed sweet potato and gluten-free flour, for example.

[Agostinho Serrano @EducatingwithAI]

its 06/05/2026, more than 3y after all GenAI spread in the world and the ONLY voice recognition tool (bundled inside a major company chatbot) that works (for non english speakers) is whisper from @OpenAI. Let that sink in. @GeminiApp , @claudeai all do a terrible job.

[Agostinho Serrano @EducatingwithAI]

Most answers are that TPU/GPU and/or competence of research teams is the reason why. And they do make sense, you can see how different tools (office integration, memory management, search past chats) are better in Claude.

But there’s a “trend” in different anthropic models - they do answer more grounded in the user context, as the model seems more “wise”, while OA models feel more “autistic”.

I just asked both if my computer automatically - when using a hotspot - informed Netflix to use lower bandwidth.

ChatGPT 5.5 explained how to verify in windows and Mac.

Claude gave me the Mac answer only.

Both have in their memory that I own a Mac, and no windows computer.

This trickle down to a massive more useful model for productivity using Claude, for real daily work.

[Agostinho Serrano @EducatingwithAI]

@johnennis Seems familiar to me with a Molecular QM background.

Back in 1990-2000 doing chemistry in silico we worried about using interfaces to understand what the Mol.Q.Mech calculations were doing, and what their numerical results meant