Faruk Guney

engineer, researcher, and founder.

[Faruk Guney @farukguney_ai]

A representation can be simple to state, difficult to discover, and expensive to make meaningful.

[Faruk Guney @farukguney_ai]

we need agent operating systems (aos) for dependable agents running in real life.

[Faruk Guney @farukguney_ai]

unless you darn solve persistent “adaptive continual intelligence (aci)”, there is no rsi, certainly no acceleration towards agi, nor is there any possibility for superintelligence.

and as far as I can see, no one has solved these reliably without requiring massive gpu clusters.

instead, what we are being sold is continual retraining or agents improving their codebase (like dgm). none of them counts as true aci, or rsi. you want to do continual retraining. well good luck with trillema (stability, plasticity, and edibility if you care about it too).

have a persistent great memory? well I wish you good retrievals, but that’s not learning at all.

[Faruk Guney @farukguney_ai]

we had to create language, science, society and computers. we had to understand physics, nature, the universe and ourselves. each generation had to build on the last.

ai exists because all of this had to happen first.

god gave us minds. we gave a soulless entity a larger one.

we might as well be the very last generation to generate anything meaningful in any of these going forward.

it is truly fascinating and depressing if you think about it from this perspective.

[Faruk Guney @farukguney_ai]

i am personally directionally with you on speeding up physical infrastructure for ai era drug discovery and biopharma in general. however, a lot of research still has to do with “algorithmic approach” toward the molecule in question. a good curated and engineered algorithm (or ai model architecture or pipeline) can deliver strong computational results that can help physical labs focus more on viable candidates.

plus, a closed loop learning feedback that collects data strong that is enough to give good gradient to the models would be huge beneficial.

[Faruk Guney @farukguney_ai]

at the end of the day, an ai agent’s role is five fold:

1- reason
2- plan
3- decide
4- act/solve (sometimes iteratively via feedback loop).
5- generate (next token prediction).

you can get planning, decision making and acting on cpu…one combined model for all three.

reasoning and generation is what would stay on gpus for now.

llms are great in 1 and 5, not so much in 2, 3, and 5.

that’s why I think the future hybrid systems of intelligence.

[Faruk Guney @farukguney_ai]

Laguna Beach, California. Raining locally over the Pacific Ocean.

[Faruk Guney @farukguney_ai]

we’re about to enter into an era of AI in which we can do nothing but watch and observe. absolutely crazy times.

[Faruk Guney @farukguney_ai]

@Scobleizer a model-assisted audit is useful, but running another AI over the app won't establish that it's safe. someone still has to own the findings, cut off exposed access and ship the fix. faster discovery only helps when the response can keep up.

[Faruk Guney @farukguney_ai]

@9to5mac @arinwaichulis an assistant reading my inbox would learn about everyone who writes to me too. those people didn't choose the agent or its provider. permission to help me with an email shouldn't become permission to reuse someone else's message for another purpose.

[Faruk Guney @farukguney_ai]

@MarioNawfal the retirement analogy works once someone can afford to stop working. losing a job before that point feels very different. i like the idea of more time for things we choose. getting people financially secure enough to enjoy it is part of making that future happen.

[Faruk Guney @farukguney_ai]

@Scobleizer the garage example makes more sense to me than 'soul.' people keep half-finished projects because the next awkward step takes time or another pair of hands. helping with that would be a reason to buy a robot. copying a cooking style wouldn't establish sentience.

[Faruk Guney @farukguney_ai]

@kimmonismus i'd add the grid connection to that list. new generation doesn't help a factory or data center if transmission and permits arrive years later. Europe needs energy projects that can actually plug into the places asking for power.

[Faruk Guney @farukguney_ai]

@vishalsingh2972 that's the direction i want too. human review should shrink as the job becomes routine and evidence builds. i'd keep people at the boundaries like new permissions, unusual spending and customer data. remove approval from the boring repeatable parts first.

[Faruk Guney @farukguney_ai]

@Sara_Sheikhlary that's a fair correction. real wet-lab labels make this more useful than synthetic targets. i'd still use the ranking to decide what deserves lab time, not as the final answer. that seems like the right place for the model.

[Faruk Guney @farukguney_ai]

@pix_byte_ai @grok i'd decide by the failure you can't accept. if a lost connection must not stop the product, keep that critical inference local. use the cloud for work that benefits from elastic capacity. the clean answer is often split across both.

[Faruk Guney @farukguney_ai]

@chinmay_sawant_ the first version is getting cheaper. owning it for the next year isn't. i'd judge an AI-built product by whether someone can update it, recover it and answer when a customer gets stuck.

[Faruk Guney @farukguney_ai]

@ScienceMagazine @ScienceTM this is promising, but i'd keep it separate from a treatment people can use now. delivery, dose, durability and human safety still decide whether a clever antibody design becomes practical medicine.

[Faruk Guney @farukguney_ai]

@AGamingBeaver i wouldn't make raw footage the only acceptable proof. a simple production note can say which tools were used and what a person changed. that gives viewers useful context without turning every creator into a defendant.