Nathan Martell

Founder & Lead Software Engineer at @Bifrost_Inc |
Building @Vivaion 🔱

[Armani Ferrante @armaniferrante]

I keep seeing terrible advice online for people interested in computer science.

AI is making programming knowledge obsolete in the same way interpreted languages made compiled languages obsolete, in the same way compilers made assembly obsoletely, and in the same way assembly made knowledge of circuits obsolete.

It didn't.

The percentage of people that can go through every layer of the stack will shrink and that knowledge will become more valuable, not less.

With more tools, your knowledge base should expand not contract. Learn everything you can, while you're in school--and use AI to accelerate that process. Learn how programming languages work. Learn how computers work. Learn how matrix multiplication is optimized. Learn FFTs are implemented. Learn calculus and linear algebra. Learn optimization. Learn how that all applies to training. Learn how to use the AI tools, but also learn how they are built.

The history of computing is the history of abstraction layers being built one on top of the other, and it's only when you can traverse layers of abstraction can you truly build new things.

Learn more. Build more.

[Solana @solana]

Solana 🤝 Fidelity

- Trade SOL direct from a Fidelity brokerage account
- Trade the Fidelity Solana ETF from any brokerage
- ???

[Nathan Martell @BifrostTitan]

@blockiosaurus Try increasing the temperature, adjusting the TopP and increase the repeat penalty parameters. If its still fried then you over-cooked🍳

[Nathan Martell @BifrostTitan]

@ilmoi odin fun runs on ICP and its like pump fun but for wrapped btc. only reason I know about internet computer.

its not that bad either. it can run LLM inference directly on-chain which is something Solana might not ever be able to do imo

[Nathan Martell @BifrostTitan]

@mert @chainyoda @MohamedFFouda The feds use a cracked out version of bubble maps with timing analysis you are not hiding shit from them. They use a combination of witnesses & CEX data to put you in prison. They will slap you with tax evasion just for wasting their time

[Nathan Martell @BifrostTitan]

Solnet v8.5.0 is available on nuget for anyone looking to 1-shot a mobile app using c# MAUI & grok

[chase @therealchaseeb]

Hello.

I'm once again telling you that if you are building for the Colosseum Cypherpunk Hackathon, then add mobile to your project.

@solanamobile is selecting up to 10 teams from the winners pool for:

- $10,000
- Marketing and Launch Support
- dApp Store Placement
- and MORE

[Nathan Martell @BifrostTitan]

What if you could see every possible outcome all at once?

The result of 7 years of research at @Bifrost_Inc showcasing 1500 machine learning forecasts over the course of 24 hours against the price of SOL 🪸

dm me if you want private access. open to partnerships as well

[NASA's Kennedy Space Center @NASAKennedy]

🖥️Live Now! Tune in for launch coverage for IMAP, SWFO-L1, and Carruthers!

Teams are on console at SpaceX's Hangar X and LSP’s Hangar AE at Cape Canaveral Space Force Station.

Where are you watching the launch?https://t.co/c2139itpbg

[Nathan Martell @BifrostTitan]

@supermarioblock its like feeding mosquitoes. your users are only going to be malicious or from the black market. I personally think you guys are talented enough to work on something other than enabling terrorism and fraud 🫣

[Nathan Martell @BifrostTitan]

github repo:
https://t.co/wIYYrIWdHX

[Nathan Martell @BifrostTitan]

Introducing Prometheus, an open-source hybrid AI assistant platform for Solana developers.

Built to generate production-ready code in seconds, it fuses cloud LLMs, local LLMs, and Semantic Kernel into a reproducible pipeline for:
🧠 Conversational codegen
📦 Structured Solana/Anchor blueprints
🔁 Auditable prompt workflows
📊 Dataset generation for LLM tuning
🦾 Custom Solana LLM training

[Solana @solana]

Bitcoin 🤝 Solana

[robin @zebird0]

Biggest challenge for solo founders isn't skillset imo

It's psychological.

Climbing Everest by yourself is challenging and much more lonely. There's no one to lean on who's as committed as you.

Heavy is the head that wears the crown

[Nathan Martell @BifrostTitan]

full process is a bit lengthy but you would create a dataset you want to train then train a open source model on it so it becomes "fine-tuned" for your downstream purpose then craft a system prompt using markdown format to encode strict rules & embed examples for in-context training.

Reversed engineered system prompts:
https://t.co/S05XMLkqKm

[Nathan Martell @BifrostTitan]

Human as judge or llm as judge?
this is what I was referring to.
https://t.co/BwVbvvaF1M

the big AI training data doesn't contain anything swap related most likely and they will "hallucinate" dependencies, functions, properties etc because it was never trained on it. those are the closest tokens that make sense in its current context

[John Carmack @ID_AA_Carmack]

Natural conversation includes interruptions and talking over people, which is hard for an LLM to model as a single autoregressive sequence. I’m sure you can get pretty far by creating a text sequence with movie-script like breaks mid sentence, but it seems like the real solution would involve parallel streams of listening and thinking with talking queued up for pauses or rising to an interruption priority. Intermixing tokens from different streams and doing something custom with attention seems plausible.

[MagicBlock ✨ @magicblock]

Any sufficiently advanced technology is indistinguishable from Magic.

[jack @jack]

goose + qwen3-coder = wow

[Bifrost @Bifrost_Inc]

Our official beta release of Mirage One is on the Microsoft store. Mirage is an AI assistant that runs locally on Windows PCs.🪟

Users can select different models trained by us while also applying a unique persona for Mirage to inherit.
🤠👽👩‍💻

https://t.co/anWkX5UFDd