Hacker News: MediaSquirrel

Scalable GANs with Transformers

MediaSquirrel — Tue, 30 Jun 2026 22:41:55 +0000

Article URL: https://arxiv.org/abs/2509.24935

Comments URL: https://news.ycombinator.com/item?id=48740210

Points: 3

# Comments: 0

Cheaper Than Concrete: Robots and the New Stone Age

MediaSquirrel — Tue, 30 Jun 2026 20:02:29 +0000

Article URL: https://originals.is/p/cheap-splendor-robots-and-the-new

Comments URL: https://news.ycombinator.com/item?id=48738449

Points: 3

# Comments: 0

Lift4D: Harmonizing Single-View 3D Estimation for 4D Reconstruction In-the-Wild

MediaSquirrel — Tue, 23 Jun 2026 20:08:18 +0000

Article URL: https://arxiv.org/abs/2606.23688

Comments URL: https://news.ycombinator.com/item?id=48650597

Points: 3

# Comments: 0

New comment by MediaSquirrel in "Death Is an Engineering Problem"

MediaSquirrel — Tue, 23 Jun 2026 19:41:11 +0000

Dr. Alex Mathiasen, PhD wants to build a pause button for human biology – so that you can live forever. All he needs is enough GPUs to simulate quantum physics.

Death Is an Engineering Problem

MediaSquirrel — Tue, 23 Jun 2026 19:41:11 +0000

Article URL: https://originals.is/p/death-is-an-engineering-problem

Comments URL: https://news.ycombinator.com/item?id=48650262

Points: 3

# Comments: 1

Non-frontal face recognition using GANs and memristor-based classifiers

MediaSquirrel — Wed, 17 Jun 2026 06:41:35 +0000

Article URL: https://arxiv.org/abs/2606.12074

Comments URL: https://news.ycombinator.com/item?id=48566644

Points: 2

# Comments: 0

MambAdapter: Lightweight Mamba-Based Adapters for Transfer Learning

MediaSquirrel — Tue, 16 Jun 2026 19:08:02 +0000

Article URL: https://arxiv.org/abs/2606.15638

Comments URL: https://news.ycombinator.com/item?id=48560367

Points: 2

# Comments: 0

Can I Buy Your KV Cache?

MediaSquirrel — Fri, 12 Jun 2026 20:14:27 +0000

Article URL: https://arxiv.org/abs/2606.13361

Comments URL: https://news.ycombinator.com/item?id=48508949

Points: 36

# Comments: 28

Before You Think: System 0, AI-Mediated Cognition and Cognitive Colonization

MediaSquirrel — Fri, 12 Jun 2026 20:12:34 +0000

Article URL: https://arxiv.org/abs/2606.13658

Comments URL: https://news.ycombinator.com/item?id=48508924

Points: 1

# Comments: 0

Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Reasoning

MediaSquirrel — Fri, 12 Jun 2026 20:10:05 +0000

Article URL: https://arxiv.org/abs/2606.13607

Comments URL: https://news.ycombinator.com/item?id=48508899

Points: 1

# Comments: 0

New comment by MediaSquirrel in "The University in the AI Era"

MediaSquirrel — Thu, 11 Jun 2026 20:50:14 +0000

Do share.

Superficial Beliefs in LLM Decision-Making

MediaSquirrel — Thu, 11 Jun 2026 20:46:33 +0000

Article URL: https://arxiv.org/abs/2606.11016

Comments URL: https://news.ycombinator.com/item?id=48496180

Points: 3

# Comments: 0

Show HN: Magenta Real-Time Music Generation Locally on iPhone, Without the GPU

MediaSquirrel — Wed, 10 Jun 2026 22:22:24 +0000

Last Thursday, Deepmind released Magenta Realtime 2 , an open source music generation model. They said it could run on Mac, but not iPhone.

As a v̵i̵b̵e̵ ̵c̵o̵d̵i̵n̵g̵ ̵a̵d̵d̵i̵c̵t̵ agentic AI maxxi and person who has melted iPhones before (link at bottom), I took that as a personal challenge and made it my weekend project.

On Saturday, I got it to run for 10min straight on an iPhone 12 Pro from 2020 without melting the phone or - shockingly - touching the GPU.

How? I chopped the model up into 5 pieces and set them each to run on different parts of Apple's system on a chip (SoC).

My past experience taught me that if you can actually leverage it, the iPhone's NPU is incredibly powerful, and power efficient. If you're doing sustained real-time generation for long periods of time on a device without a fan, you gotta use the neural engine or else you will melt the device.

See: https://accelerateordie.com/p/we-melted-iphones-for-science

The Apple Neural Engine has a ton of constraints, the main one being that it only accepts fixed shape inputs, and only supports some architectures -- which is why I chopped the model up into pieces.

But it works! And I wrote zero lines of code by hand. Back when I was running VC-backed companies, I would have needed a small team of grumpy greybeard engineers to do this and it would have taken 2-6 weeks. Now I can feed my own nerd fetish and do this stuff myself.

Next up: I'm building an iPhone app that ties into your heart rate, movement data, location etc to generate a real-time soundtrack to you life.

What a time to be alive!

update: Demo video of it running on my iPhone 15 Pro: https://x.com/mattmireles/status/2064849702156415214?s=46

Comments URL: https://news.ycombinator.com/item?id=48483562

Points: 9

# Comments: 0

Feedback Alignment in Self-Distillation

MediaSquirrel — Wed, 10 Jun 2026 22:13:12 +0000

Article URL: https://arxiv.org/abs/2606.11173

Comments URL: https://news.ycombinator.com/item?id=48483451

Points: 2

# Comments: 0

WWDC 2026 – On-Device AI Deep Dive

MediaSquirrel — Wed, 10 Jun 2026 22:09:45 +0000

Article URL: https://gist.is/docs.google.com/en/deqIp-AK6Oxc

Comments URL: https://news.ycombinator.com/item?id=48483410

Points: 1

# Comments: 0

New comment by MediaSquirrel in "Show HN: Gemma 4 Multimodal Fine-Tuner for Apple Silicon"

MediaSquirrel — Wed, 08 Apr 2026 04:46:07 +0000

yeah, it came out after I stared on my project last year. Only issue is that you can't fine-tune it on Apple Silicon.

New comment by MediaSquirrel in "Show HN: Gemma 4 Multimodal Fine-Tuner for Apple Silicon"

MediaSquirrel — Wed, 08 Apr 2026 04:43:16 +0000

depends on the model!

If you run a smaller whisper-distil variant AND you optimize the decoder to run on Apple Neural Engine, you can get latency down to ~300ms without any backend infra.

The issue is that the smaller models tend to suck, which is why the fine-tuning is valuable.

My hypothesis is that you can distill a giant model like Gemini into a tiny distilled whisper model.

but it depends on the machina you are running, which is why local AI is a PITA.

New comment by MediaSquirrel in "Show HN: Gemma 4 Multimodal Fine-Tuner for Apple Silicon"

MediaSquirrel — Wed, 08 Apr 2026 02:04:59 +0000

Look inside here: https://github.com/mattmireles/gemma-tuner-multimodal/tree/m...

Here’s the trick: use Gemini Pro deep research to create “Advanced Hacker’s Field Guide for X” where X is the problem that you are trying to solve. Ask for all the known issues, common bugs, unintuitive patterns, etc. Get very detailed if you want.

Then feed that to Claude / Codex / Cursor. Basically, create a cheat sheet for your AI agents.

This will unlock a whole new level of capability.

I’m @mattmireles on Twitter — feel free to DM me.

New comment by MediaSquirrel in "Show HN: Gemma 4 Multimodal Fine-Tuner for Apple Silicon"

MediaSquirrel — Wed, 08 Apr 2026 01:57:04 +0000

More data -> better, faster on-device models

The actual plan was to distill Gemini 2.5 Pro into the best on-device voice dictation model.

Pretty sure it would have worked. Alas.

New comment by MediaSquirrel in "Show HN: Gemma 4 Multimodal Fine-Tuner for Apple Silicon"

MediaSquirrel — Wed, 08 Apr 2026 00:22:25 +0000

re: Whisper v3 -- how is this possible? Whisper has a 30s context window. You have to chunk it.