Just a Nerd wishing for nerd stuff
127.0.0.1
Joined February 2009
- Tweets502
- Following88
- Followers693
- Likes137
Server-side evidence for FPS cheats. Not another driver on the player's PC.
fpsdet scores what a dedicated server already accepted: gear rules, the best humans you have measured, and what that client could have known. It never bans. A person reviews the case.
Nothing installed on players. Plain Python. First look on logs you already have: github.com/Nimdy/fpsdet
@WARDOGS @BULKHEAD @Brammflakes @HowardPhilpott
#WARDOGS #anticheat
27
I'm about 95% done with a solo project: a fully local AI film studio. Not just music videos.
Write a whole story, and it generates a full-length video of exactly what you wrote. No cloud, no hosted APIs.
What I've enjoyed most isn't the output. It's building the architecture and R&D systems around the models.
What it does:
1. Story, song, text, image or video in → finished film out, 100% local on an NVIDIA DGX Spark (it runs on an RTX 5090 too)
2. A local LLM "Director" (Qwen3.6-35B) reads your whole story or song, keeps it as a Project Bible, and plans the film scene by scene, citing it at every step
3. FLUX.2 klein for stills and LTX 2.5 (22B) for motion, orchestrated through ComfyUI
4. Characters, places and props live in a reusable library, so the same person and the same room carry from the first shot to the last
5. Characters speak their lines with their own voices, lip-synced by LTX 2.5
6. Every render carries a receipt: hash-bound inputs, the exact prompt, the seed that actually ran. No hidden prompt text.
Why it doesn't drift:
Most AI video falls apart over length: faces change, rooms change, the story wanders. I built the architecture so it can't quietly happen:
• Identity comes from hash-bound reference images, not descriptions. A character is the picture you approved, verified on every render.
• Places carry a shared world "look", so ten rooms read as one building.
• Your words are bound to the render verbatim. Nothing gets rewritten behind your back.
• Nothing is overwritten silently: every render, choice and input is traceable back to where it came from.
Why I love the R&D side:
Generative models are black boxes that fail quietly, so I built instruments to look inside them:
- Counterfactual Prism: a custom ComfyUI node pack that taps the denoiser mid-sample and runs "what if" arms at every step, behind a parity gate that proves the taps don't change the output.
- Visualize: trace any frame, by hash, back to the words that made it.
- Slot tracer: compiles every workflow twice per input and diffs the graphs. 67 packages, 685 slots, 0 dead.
What measurement taught me that intuition didn't:
• Words can't control how far the camera moves. An image guide on the last frame can: it landed within 6% of the target.
• At CFG 1, the negative prompt does nothing at all.
• Ask for "a reflection of him" and you get a second person. The model adds; it doesn't transform.
• One full-body photo plus a camera arc gives a clean turnaround with identity intact. That became the training source for my character LoRAs, which now train on the same box through the same GPU queue.
My rule for the whole project: a number you haven't checked against the actual picture is a hypothesis. Open the images. The renders win.
Happy to talk with anyone building long-form generative video.
#GenerativeAI #LocalAI #ComfyUI #LTXVideo #AIFilmmaking #MachineLearning #BuildInPublic
@ltx_io
Okay... serious question.
Any legit sound engineers / audio engineers / DSP people on here?
I want somebody who actually lives in this world to watch this SIGNAL/3 video and tell me what they think.
I built the audio engine from the ground up.
Timing/scheduling. Synthesis. DSP. Routing. Racks. Buses. Returns. Automation. Playback. Capture. Analysis. Provenance. The whole damn stack.
It is NOT TidalCycles or Strudel hiding underneath with another UI slapped on top. It is not some existing DAW engine doing the hard work behind the curtain.
SIGNAL/3 is the engine.
And yeah... you can live code the music using shorthand AND live code the visualizations using shorthand.
You can stay at the creator level and say what you want musically/visually, or keep drilling all the way down into the actual source, notes, synthesis, routing, DSP, runtime state, rendered output, visual programs and GPU state underneath it.
In this video I can pick a moment in a record, figure out exactly what is making the sound, isolate it, hear it dry, post-rack, with returns, back in the mix, and trace it from the musical source all the way through the signal chain to the measured master output.
The funny part is... I know enough audio engineering to architect and build this platform.
I do NOT know enough of the actual sound-engineer lexicon to know which parts of this will make one of you go:
"holy shit... I pooped myself."
😂
So if audio engineering / DSP / production is your world, hit me up.
I want people who actually know this shit to beat on SIGNAL/3 before I call the baseline solid.
Not looking for somebody to make a track for me.
I want experts to look at the actual instrument, engine and workflow and tell me what I got right, what I got wrong, and what you would want if you were using something like this professionally.
Also... if somebody randomly knows Dr. Dre or Jimmy Iovine, send them this shit. 😂
Their recent conversation about AI becoming another creative tool in the studio is exactly the kind of future I have been thinking about while building this.
AI-native.
AI-optional.
Same instrument for the human and the AI.
No hidden magic path.
#Signal3 #AudioEngineering #SoundEngineering #DSP #MusicProduction #MusicTech #CreativeCoding #LiveCoding #WebAudio #WebGPU #GenerativeArt #CreativeAI #ArtificialIntelligence #OpenSource #Synthesizer #AudioDSP #MusicTechnology @elonmusk @drdre @suno
82
What’s the actual plan here? #Bots #AISpam #CyberSecurity #AI #BotApocalypseWe spent years blocking bots…
Now everybody is creating them. Result: more spam, more malicious bots, more marketing harvesting, more WTF.
@elonmusk @grok @briankrebs @troyhunt @SwiftOnSecurity
1
1
81
I just released JARVIS Oracle Edition as an open-source project.
This is not a chatbot wrapper.
It is a local-first cognitive architecture I have been building around persistent memory, perception, self-reflection, truth calibration, and human-governed self-improvement.
The system runs as a two-device architecture:
Raspberry Pi 5 + AI hardware as the sensor layer
Desktop GPU machine as the brain
The goal is not to pretend this is AGI.
The goal is to build the foundation for personal AI systems that can live with a person over time, remember responsibly, perceive the environment, reason about their own state, and improve under strict evidence and safety gates.
Some of what is already in the project:
- persistent memory and episodic recall
- local LLM reasoning
- voice, face, scene, and environment perception
- Raspberry Pi sensor integration
- truth calibration
- capability boundaries
- self-reflection and introspection
- neural policy layers
- human-approved self-improvement
- restart-honest continuity
- epistemic guardrails to prevent fake capability claims
One of the most important design principles is simple:
The system is not allowed to claim it can do something just because the language model can say it.
1. Capabilities have to be verified.
2. Memory has to be grounded.
3. Self-reports have to trace back to real subsystem data.
4. New authority has to be earned through evidence.
This release is the first public foundation release. It is not polished consumer software. It is a research/workshop architecture for builders who care about local AI, edge perception, memory systems, cognitive architectures, and the hard engineering behind systems that are supposed to grow over time.
No hype.
No fake magic.
Build, break, trace, improve.
@elonmusk @AnthropicAI @OpenAI @GeminiApp
Repo: github.com/Nimdy/jarvis-orac…
#LocalAI #ArtificialIntelligence #EdgeAI #RaspberryPi #LLM #OpenSource #AIAgents #CognitiveArchitecture #MachineLearning #SelfImprovingSystems
1
66
Elon, the world desperately needs trustworthy voting.
@Nimdy built github.com/Nimdy/Verifiable-… — an open-source, end-to-end verifiable voting system where the trust root is battle-tested cryptography (not blockchain hype).
✅ Cast-as-intended, recorded-as-cast, counted-as-recorded
✅ Public transcripts anyone can independently verify (two cross-checked verifiers in TS + Python)
✅ Supports clubs, HOAs, DAOs, boards, shareholder votes — with ranked choice, IRV, multi-contest ballots
✅ Paper + risk-limiting audit hybrid for real elections
✅ Coercion mitigation, everlasting privacy, zero-knowledge proofs
No more "trust us" black boxes. Every attack in the playground gets caught. Live demo: nimdy.github.io/Verifiable-V…
This is what transparent governance should look like. Let's make verifiable elections the norm. Who's in?
#VerifiableVoting #ElectionIntegrity #OpenSource #Crypto #DAO
4
53
Meet Zero and Bandwidth.
Not just faces in a video. Not just characters passing through a song.
Zero carries the weight. Bandwidth carries the signal.
Together, they live in the hidden meeting between music, image, story, and you.
And they are not done yet.
#ZeroBandwidth #AIActors #AIFilmmaking
🤖 Made with AI
71
Just shipped SSTDREAM Visual builder for SST v4 & Vercel.
Draw your full-stack app → Simulate wiring → Export clean, verified IaC you deploy yourself.
28 SST kinds + 22 Vercel kinds. Zero credentials. Zero hallucinations. 100% in-browser.
Live demo: solonerds.github.io/sstdream…
Repo: github.com/SoloNerds/sstdrea…
Perfect for solo founders tired of IaC pain.
#SST #Vercel #AWS #NextJS #IaC #WebDev
1
81
Honest scope: Perfect for organizations, DAOs, and communities. Not for high-stakes government remote voting (endpoint + coercion problems remain — as experts correctly warn).
But for everything else? This could be production-ready with more eyes and audits.
Star it, fork it, review the crypto: github.com/Nimdy/Verifiable-…
16
Why this matters:
Most "blockchain voting" projects ignore 40+ years of election security research and create new attack surfaces.
This one learns from Helios, Belenios, ElectionGuard, etc. — but goes further with:
- Verifiable mixnets for IRV
- Homomorphic tallies
- Pedersen commitments for post-quantum privacy
- Signed Merkle roots (optional chain anchor, never the trust root)
17
I just shipped 12 interactive chapters covering 80+ years of AI history.
From the 1943 Perceptron to 2024 State Space Models.
Not videos. Not slideshows.
You train the actual algorithms. In your browser. From scratch.
What you experience:
- Watch a perceptron fail at XOR, then add hidden layers
- See RNN gradients vanish, then switch to LSTM gates
- Battle a GAN's generator vs discriminator on live data
- Dissolve data into noise and reverse it (Diffusion Models)
- Rank responses to train a reward model (RLHF)
- Grow a model and watch emergent abilities turn on (Scaling Laws)
- Compare O(n) SSMs vs O(n^2) Transformers in real time
Every implementation is hand-written from scratch. No library calls. Real math. Real gradients.
Zero setup. Zero cost. Everything runs locally.
Try it: hackme.ai
Like what I am building? If so, reach out and lets do some nerd sessions and build other things!
#AI #MachineLearning #DeepLearning #NeuralNetworks #BuildInPublic
🤖 Made with AI
27
youtube.com/watch?v=FP77IrZE…
@ParamountPics lets go - give this man some runway and let him create the new 100M+ opening night for a real TMNT movie!
#teenagemutantNinjaturtles
46