frontpage.
newsnewestaskshowjobs

Made with ♥ by @iamnishanth

Open Source @Github

fp.

Open in hackernews

Show HN: Ghost Engine – generate weights on the fly

https://github.com/sajanlamsal/ghost-engine
1•saznlamsal•1h ago
Hello HN, I’m the author of Ghost Engine.

I built this to challenge the assumption that we are strictly bound by the "Memory Wall." My hypothesis was that modern consumer silicon (like Apple M-series) has enough spare compute to decompress weights procedurally faster than it can read them from RAM.

The Architecture: It uses a "Predator-Prey" method:

Predators: We identify and preserve the high-magnitude outliers (the "Alpha" weights) in FP16.

Prey: The remaining weights are compressed into ternary masks {-1, 0, 1} and a block-wise scalar.

Reconstruction: A bitwise kernel reconstructs the layer in L2 cache during the forward pass.

Results: I validated this on Llama-3-8B (Layer 20, SwiGLU).

Compression: ~3.0 bits per weight (effective).

Fidelity: 0.915 Cosine Similarity (weights) / 0.912 (outputs).

Size: Brings an 8B model down to ~3GB.

The repo is currently a "Proof of Engine" in Python/MLX. The math works, but to realize the theoretical speed gains (125 t/s), I am working on porting the decompression kernels to Metal/CUDA.

Happy to answer questions about the compression logic or the "Predator" selection algorithm!

Comments

saznlamsal•1h ago
Hello HN, I’m the author of Ghost Engine.

I built this to challenge the assumption that we are strictly bound by the "Memory Wall." My hypothesis was that modern consumer silicon (like Apple M-series) has enough spare compute to decompress weights procedurally faster than it can read them from RAM.

The Architecture: It uses a "Predator-Prey" method:

Predators: We identify and preserve the high-magnitude outliers (the "Alpha" weights) in FP16.

Prey: The remaining weights are compressed into ternary masks {-1, 0, 1} and a block-wise scalar.

Reconstruction: A bitwise kernel reconstructs the layer in L2 cache during the forward pass.

Results: I validated this on Llama-3-8B (Layer 20, SwiGLU).

Compression: ~3.0 bits per weight (effective).

Fidelity: 0.915 Cosine Similarity (weights) / 0.912 (outputs).

Size: Brings an 8B model down to ~3GB.

The repo is currently a "Proof of Engine" in Python/MLX. The math works, but to realize the theoretical speed gains (125 t/s), I am working on porting the decompression kernels to Metal/CUDA.

Happy to answer questions about the compression logic or the "Predator" selection algorithm!

Apple testing new App Store design that blurs the line between ads and results

https://9to5mac.com/2026/01/16/iphone-apple-app-store-search-results-ads-new-design/
1•ksec•1m ago•0 comments

GoCrazyAI – AI image and video generator

1•gocrazyai•2m ago•0 comments

NASA Technology Spinoffs Archives

https://spinoff.nasa.gov/spinoff/archives
1•o4c•3m ago•0 comments

Things GenAI Needs for Better Content Design [video]

https://www.youtube.com/watch?v=WfHBINzB0Ps
1•ArneVogel•3m ago•0 comments

An idea that several novices tried to complete on a weekend

https://github.com/hdcola/metapi
1•hdcola•4m ago•0 comments

Teaching LLMs to Stop Wasting Tokens

https://codereviewr.app/blog/teach-llms-to-stop-wasting-tokens
1•sousvidal•5m ago•0 comments

Technology Readiness Levels

https://www.nasa.gov/directorates/somd/space-communications-navigation-program/technology-readine...
1•o4c•5m ago•0 comments

Stocks sell off globally as traders digest Trump message

https://fortune.com/2026/01/19/stocks-sell-off-trump-text-message-greenland-nobel-peace-prize/
1•doener•5m ago•0 comments

Removal of GTK2 from forky (Debian 14)

https://lists.debian.org/debian-devel/2026/01/msg00090.html
1•birdculture•9m ago•0 comments

Building live collaboration in Rust for users, part 1

https://www.photoroom.com/inside-photoroom/building-live-collaboration-in-rust-for-millions-of-us...
1•ea016•11m ago•0 comments

Show HN: "htop" for PyTorch training, see stalls, memory and step time live

1•traceopt•12m ago•0 comments

Show HN: Tsshd – mosh-like SSH with QUIC, roaming and full OpenSSH compatibility

https://github.com/trzsz/tsshd
1•LonnyWong•12m ago•0 comments

Generate professional App Store previews instantly with AI

https://appscreenshotstudio.com/
1•Welten01•13m ago•1 comments

Show HN: PolicyBind – AI Policy-as-Code with real-time token access control

https://github.com/clay-good/policybind
1•hireclay•14m ago•0 comments

Ask HN: Will there be resurgence of webapps instead of app store?

1•omnifischer•14m ago•1 comments

Sled is Claude Code on your mobile with voice

https://sled.layercode.com
2•dctanner•14m ago•1 comments

Happy 40th birthday Apple Lisa

https://computerhistory.org/blog/the-lisa-apples-most-influential-failure
1•stmw•16m ago•0 comments

'Are You Dead?': The viral Chinese app for young people living alone

https://www.bbc.com/news/articles/c3381r5nnn6o
2•mooreds•18m ago•0 comments

Causes of global extinctions in the history of life: facts and hypotheses (2020)

https://pmc.ncbi.nlm.nih.gov/articles/PMC7716527/
2•mooreds•18m ago•0 comments

The Memory-Transfer Episode

https://www.apa.org/monitor/2010/06/memory-transfer
1•34679•19m ago•0 comments

Understanding the psychology behind product decisions

https://designexplained.substack.com/p/understanding-the-psychology-behind
1•kaizenb•21m ago•0 comments

Mestastic on Family Cruise – Worked Great for Family of 4

https://old.reddit.com/r/meshtastic/comments/1qd2z97/mestastic_on_family_cruise_worked_great_for/
2•lormayna•22m ago•0 comments

Back-scratching bovine leads scientists to reassess intelligence of cows

https://www.theguardian.com/science/2026/jan/19/back-scratching-cow-veronika-bovine-intelligence
2•n1b0m•23m ago•0 comments

Show HN: Eigent – the open source alternative of Cowork

1•camelaiorg•24m ago•0 comments

Let's Play Quakeworld

https://fabiensanglard.net/quakeworld/
1•_pob•24m ago•0 comments

How to Get Your First Users [video]

https://www.youtube.com/watch?v=0kARDVL2nZg
1•gmays•25m ago•0 comments

2026 Linux Audio Conference Focus on LLMs and Floss

https://lac26.mucs.club/
3•Lanedo•26m ago•0 comments

Show HN: Pipenet – A Modern Alternative to Localtunnel

https://pipenet.dev/
3•punkpeye•27m ago•0 comments

PicoPCMCIA – Yyzkevin

https://www.yyzkevin.com/picopcmcia/
1•rbanffy•27m ago•0 comments

New milestones for Nyno (open-source n8n alternative for AI Workflows, Jan. 26)

https://nyno.dev/new-milestones-for-nyno-open-source-n8n-alternative-for-ai-workflows-january-2026
3•theyogadev•28m ago•2 comments