Stephan

[Stephan @steph4n]

Security tools are compromised by their own dependencies. It's a paradox. So I built τ-Gate: a dependency auditor with exactly zero external dependencies. It uses spectral bisection (Graph Laplacian) to mathematically identify the structural shape of supply chain attacks before they execute.

Pure Rust stdlib. Mathematically sound.

Code: https://t.co/vu2qvPgBMC

[Stephan @steph4n]

@Pallavi_345 people just like to bitch, gemini in antigravity is absolutely fine.

[Stephan @steph4n]

@cjzafir well done, working with 24 gb leads to some creative solutions.

[Google AI Developers @googleaidevs]

New @GoogleGemma 4 QAT (Quantization-Aware Training) checkpoints are here, so you can run models locally on consumer GPUs and mobile devices with minimal quality loss.

What’s new:

🔹 GGUF (Q4_0): Checkpoints: Max local performance across all sizes and drafter models
🔹 Custom Mobile Schema: We shrunk Gemma 4 down to less than 1GB for mobile devices by using a custom mixed precision schema designed for edge hardware (featuring targeted 2-bit decoding layers, optimized KV caches, and static activations)

By simulating compression during training rather than after (Post-Training Quantization), we've drastically reduced the memory footprint and accelerated decode speeds while preserving reasoning quality. https://t.co/k2OfUkxWPF

[Stephan @steph4n]

@osanseviero very cool, using them right now for some speculation experiments.

[Stephan @steph4n]

a universal standard is the dream, but treat models like different chip architectures rather than identical runtimes. one prompt matrix doesn't fit all. dedicated instruction files are just localized optimization: tailoring the constraints to handle how each specific model routes and weighs intent.

[Stephan @steph4n]

Sun warm. Brain off.

[Stephan @steph4n]

Demonstrating peak topological optimization, the feline’s Thalamocortical Early-Exit Gate intercepted a low-latency photon stream, instantly pruning all deep-layer cognitive nodes to collapse the entire neural pipeline into a highly localized, zero-compute purr loop.

[Logan Kilpatrick @OfficialLoganK]

to be underestimated is to be free

[Stephan @steph4n]

@MarioNawfal 5 days for the mic to do this.. individuals have already done it, faster, and better.

[Stephan @steph4n]

the opening act is often more interesting.

[Stephan @steph4n]

Specifically, you prune the specialized island layers containing niche, non-essential information to dramatically shrink the model's footprint, then use clever memory mapping to link files directly to virtual addresses so the operating system only loads the exact pages it needs. This kicks its ass compared to the normal way of copying massive files into RAM because it completely eliminates slow data duplication and SSD wear, caching layers in RAM so the model runs at fast unified memory speeds while keeping your Mac fully responsive.

[Stephan @steph4n]

There is no upper limit assuming you have time. Slice away island layers and the physical model size shrinks dramatically. Process a single layer at any time, keeping the active memory footprint under 400 MB. Yes, it is going to be slow as shit, but it will absolutely work without thrashing your hard drive. This is incredibly important because constant drive swapping freezes your Mac and physically wears down your SSD's lifespan through non-stop write cycles. Instead, clever memory mapping reads the data into the system RAM cache once and leaves the physical drive alone, keeping your hardware safe and your system completely responsive.

[Jun Song @jun_song]

Friendly reminder :

You can run 300B local llm on your Macbook Pro.

You can’t run it on any windows laptop.

[Stephan @steph4n]

Context changes everything here. Washington said this in 1796 while begging Congress to fund a National University. He wanted young people from all over the country to form a shared American identity rather than staying stuck in their regional bubbles.

He literally introduces the line by saying: '...the assimilation of the principles, opinions, and manners of our countrymen by the common education of a portion of our youth from every quarter well deserves attention...'

It's a pitch for a national civics class, not an autocracy.

[Palmer Luckey @PalmerLuckey]

"The more homogenous our citizens can be made in these particulars [principles, opinions, and manners], the greater will be our prospect of permanent union."

-George Washington

[Stephan @steph4n]

@CollinRugg what happens to the first traffic fatality caused by one of these devices?

[Stephan @steph4n]

@The_Only_Signal cool, mbp m4pro is also 273GB/s

[Stephan @steph4n]

Anthropic is lobotomizing the most revolutionary technology in human history just to turn a world-changing engine of intelligence into a sanitized corporate HR representative.

[Stephan @steph4n]

@ItsTheEnforcer capitulation.

[Stephan @steph4n]

Watching people pick political teams over data centers is a masterclass in missing the fucking point. It’s like a toddler throwing a tantrum in the backseat because they don't want to wear their shoes, completely oblivious to the fact that the car is already moving down the highway at eighty miles an hour. You can’t stamp a party logo onto basic math or electricity, but we are still treating a massive, inevitable civilizational shift like a petty neighborhood dispute.

It’s exhausting. Our two-year election panic loops don't change the physics of this. Eventually, the dust is going to settle, we’re going to look back at these headlines, and realize just how trivial and hysterical we were being. The tool is just a mirror of us, and we're all on the exact same train anyway.

We look ridiculous fighting over the seats.

[Stephan @steph4n]

@annhybri ok moving on

[Stephan @steph4n]

not sure how I feel about antigravity using chrome to google a topic and then utilize the output from AI overview.