Balázs

kz/hu in 🇩🇰; econ & cs @AarhusUni, @AaltoUniversity; fellow @fastdotai; expert @kaggle; learning and building just for the love of the game

[Balázs @balazsfab]

@kyliebytes @ashleevance @corememory Can you tell me where is it? Just to make sure I will avoid that area

[Balázs @balazsfab]

The real moat is being asexual.

[Balázs @balazsfab]

AlphaGenome 2 precomputing scores for 9 billion variants turns variant effect prediction into a lookup, so the edge moves to labels and cell context. https://t.co/S9Mxrt930S

[Demis Hassabis @demishassabis]

Super proud to announce AlphaGenome 2 - our new DNA sequence model! 🧬 It predicts the impact of DNA variants in different specific cell types - unlocking crucial information for understanding complex diseases. Available for academic research via our API: https://t.co/dEyrvpTxIl

[Balázs @balazsfab]

@Polymarket Low key that would make him the Trump of Silicon Valley.

[Balázs @balazsfab]

Fable 5.2 or Gemini 4 Carbon, which one would you trust to write your validation split without checking its work?

[Balázs @balazsfab]

@elliotarledge If Carbon's curve was still rising at 2.5 hours, the 23.8x vs Opus 35.5x gap looks more like budget than skill.

[Balázs @balazsfab]

@MParakhin Carbon beating Fable 5.2 on math at only High effort but not on code is odd, is it correctness or code style that breaks?

[Balázs @balazsfab]

@BenBajarin A 6 point yearly climb from 33.8% to 45.6% puts AMD over 50% of server revenue around mid-2027.

[Balázs @balazsfab]

@BenBajarin Apple at 11x earnings in 2015 is the reminder that the market prices durability last, usually years too late.

[Balázs @balazsfab]

If CPUs really go from $62B to $251B by 2030, DPUs could quietly be the biggest winner of the agent wave. https://t.co/4JuXnFQREm

[The Diligence Stack @DiligenceStack]

Agentic CPU racks sit outside the GPU scale-up domain, so they create a separate scale-out network. Add untrusted agents and DPUs become the isolation and control layer for that new CPU fabric. https://t.co/WmQ6oaoLID

[Balázs @balazsfab]

@elliotarledge My bet is Gemini 4 Carbon gets correctness right but rarely beats torch.compile on speed, curious to see the numbers.

[Balázs @balazsfab]

@FrancoisChauba1 Local wins most daily tasks, but long context stays in the cloud for a while since phones can't hold huge KV caches.

[Balázs @balazsfab]

@paulg Swapping a central post for 7 ball bearings is the kind of fix that only looks obvious once Rolex copies it.

[Balázs @balazsfab]

@Hesamation Did Fable 5.2 get close to beating your own net worth score, like Sol 6.1 did?

[Balázs @balazsfab]

@BenBajarin Pooled unified memory with no command line makes a few Macs over Thunderbolt 5 the easiest private LLM cluster you can buy.

[Balázs @balazsfab]

@OliverMolander No hobbies is a burnout timeline, and plenty of great startups began as someone's side interest.

[Balázs @balazsfab]

@ekzhang1 KV storage is the sleeper, once cache hits are cheap, prompt caching becomes a pricing lever, not just a latency trick.

[Balázs @balazsfab]

@can Better at bugs and style for sure, but only a human asks whether the PR should exist at all.

[Balázs @balazsfab]

Pooled unified memory with no command line makes a few Macs over Thunderbolt 5 the easiest private LLM cluster you can buy Fter this.

[Aaron @aaronp613]

Apple is working on a new app called Harbor

Harbor turns your Mac into a private AI inference server. Run open-weight models locally on Apple silicon—your data never leaves your hardware, and nothing reaches the cloud. Connect multiple Mac computers over Thunderbolt 5 to pool unified memory, enabling large models and higher concurrency that a single computer can handle alone.

Set up in minutes, not hours Harbor automatically detects your devices over Thunderbolt 5 to form a cluster. Plug in, start, and you’re running—no manual configuration, no command line, no infrastructure expertise required.

Download and run models directly Browse, download, and load open-weight models from within the app. Harbor handles model management so you can focus on the work, not the setup.

Cluster for scale Connect two or more Thunderbolt 5-capable Mac computers and Harbor pools their unified memory into a single inference resource. Run models that exceed any single computer’s capacity, or serve more concurrent requests across your cluster.

Harbor is currently in a private preview under the Claris brand.

[Matias @Bolmercl]

@k0csis @Hesamation Bro you are so close to being enlightened from first principles lmao