Nash A. | That Upgrade Feeling

316 posts

Nash A. | That Upgrade Feeling

@UpgradeFeeling

Engineer & founder. I explore AI, software & tech to find what actually feels like an upgrade.

Worldwide Katılım Nisan 2012

336 Takip Edilen46 Takipçiler

Sabitlenmiş Tweet

Nash A. | That Upgrade Feeling@UpgradeFeeling·3d

Is buiding your own AI only possible for giant companies? After looking into it, I found the answer is no. You can build a useful AI without training a massive model from scratch or spending millions. But first, what does "building AI" actually mean? 1/31

English

239

Nash A. | That Upgrade Feeling@UpgradeFeeling·18m

25/31 But training a frontier model from zero is a completely different game. You would need massive datasets and the legal rights to use them, thousands of GPUs, storage, networking, researchers, engineers, safety systems, evaluation, and many failed experiments. This is still mainly possible for large companies and research labs.

English

Nash A. | That Upgrade Feeling@UpgradeFeeling·3d

English

239

Nash A. | That Upgrade Feeling@UpgradeFeeling·34m

24/31 And AI is not only about chatbots. You can also train smaller models from zero for things like: • Prediction • Classification • Fault detection • Recommendations • Image analysis • Anomaly detection • Demand forecasting Some of these models can run on a normal computer and cost almost nothing.

English

Nash A. | That Upgrade Feeling@UpgradeFeeling·2h

23/31 A rough example: 8 A100 GPUs × 100 hours × $1.39 per hour = about $1,112 And this is only the compute cost. The real project may cost much more after data preparation, storage, failed experiments, evaluation, and engineering time.

English

Nash A. | That Upgrade Feeling@UpgradeFeeling·3h

22/31 There is also something called continued pretraining. You take an open model and continue training it using a large amount of specialized text before fine-tuning it. This can give the model deeper knowledge in one domain, but it also needs more data and much more compute.

English

Nash A. | That Upgrade Feeling@UpgradeFeeling·4h

I’m really amazed by what new technology has brought us. It makes studying much easier, less boring, and more interactive because you can actually interact with the 3D models.

The Bugged Dev@thebuggeddev

Only if education could be this interactive ❤️‍🔥 I've had a looong wish to build something genuinely useful through vibe coding, and I finally did it. A 3D human anatomy application built with @threejs using GPT 5.6 Sol. It all started with a single design image that I created using GPT Image 2.0. I then used it to generate every 3D organ image, one by one. Next, I converted each of those images into 3D models using @tripoai (and no, they didn't sponsor this 😄). After that, I opened Codex, wrote a master prompt based on the design, and gave it the prompt, the design image, and all the 3D models. Codex built the first version beautifully, but there was one big problem. Every single 3D model was nearly 120-150 MB. That obviously wasn't practical for the web and was giving a performance of 16fps. After a few iterations, Codex optimized each model down to roughly 2–5.5 MB while preserving the visual quality, reducing the total asset size from ~900 MB to just 28.6 MB. And each model loads on demand. Along the way, Codex also generated those anatomical illustrations showing where each organ sits in the human body, and even created the interactive hotspot markers that explain different parts of every organ. It handled all of that. The process wasn't exactly one shot, but it also wasn't difficult. You just have to do it step by step. It genuinely felt like building something that could make learning anatomy much more engaging. The inspiration came from @DilumSanjaya's 3D animal plant cell project. I remember seeing it and thinking, "I want to build something like this one day." And I did it :D Live: anatomy-livid.vercel.app Code: github.com/thebuggeddev/a…

English

Nash A. | That Upgrade Feeling@UpgradeFeeling·4h

No, I don’t review the code line by line or go behind the AI. I developed my own way by splitting the project into a roadmap, adding a vibe coding policy on top of it, and letting the AI test every stage through PowerShell before moving to the next one. But I don’t recommend anyone do this. I only use this method for my personal projects.

English

Seb@plainionist·14h

Do you still review AI generated Code? 🤔

English

211

112

19.7K

Nash A. | That Upgrade Feeling@UpgradeFeeling·4h

21/31 The seventh way is distillation. Here, you use a powerful model as a teacher. The powerful model generates examples, and then you train a smaller model to copy that behavior. The goal is to create a model that is: • Smaller • Faster • Cheaper to run • Specialized for one task

English

Nash A. | That Upgrade Feeling@UpgradeFeeling·5h

You are kind of right when it comes to recovery from long relationships, but I think six months is enough. If things still can’t be fixed, then you should move on. Don’t keep trying to fix a dead relationship. It is like watching the same movie twice when you already know how it ends.

English

Unkonfined@unkonfined·5h

Everyone should be forced to remain single for one year after a relationship they’re in ends. Being single is glorious and good for you.

English

276

153

700

20.1K

Nash A. | That Upgrade Feeling@UpgradeFeeling·5h

20/31 Also, open-weight does not always mean the model is completely open source. Some models allow you to download the weights, but they still have license restrictions. So before using one commercially, you should read the license carefully. Free to download does not always mean free for every business use.

English

Nash A. | That Upgrade Feeling@UpgradeFeeling·6h

@NanouuSymeon It depends on the lighting in the room. If the room is dark, I prefer dark mode because it reduces glare. But when the room is bright, I prefer light mode.

English

277

• nanou •@NanouuSymeon·6h

As a developer, do you prefer dark mode or light mode?

English

139

102

9.4K

Nash A. | That Upgrade Feeling@UpgradeFeeling·6h

19/31 But running the model privately does not automatically mean it will be cheaper. At those hourly rates, a cloud RTX 4090 running nonstop for 30 days would cost around $497. An A100 would cost around $1,001. For light usage, I think an API may actually be cheaper. Private hosting makes more sense when privacy, control, or scale justifies the cost.

English

Nash A. | That Upgrade Feeling@UpgradeFeeling·8h

2/8 A model, an app and an agent are not actually the same thing. The model is basically the brain behind the system. The app is the way you communicate with that model. But the agent can use the model, tools and information together to complete something in many steps.

Nash A. | That Upgrade Feeling tweet media

English

Nash A. | That Upgrade Feeling@UpgradeFeeling·13h

People these days keep talking about AI agents, and this let me think what an AI agent actually is. The name sounds complicated, but I think the idea is really simple. An AI model gives you an answer. But an AI agent can take a goal, make decisions, use tools and actually do something for you. 1/8

English

Nash A. | That Upgrade Feeling@UpgradeFeeling·8h

18/31 The sixth way is running an open-weight model yourself. You can run it: • On your computer • On a private server • On a cloud GPU • Inside a company network This gives you more privacy, more control, and more freedom to customize the system.

English

Nash A. | That Upgrade Feeling@UpgradeFeeling·14h

17/31 But obviously, compute is not the full cost. You still need to prepare and clean the data, run experiments, evaluate the results, store the model, and sometimes repeat the training many times. I think in many AI projects, preparing good data takes more work than using the GPU itself.

English

Nash A. | That Upgrade Feeling@UpgradeFeeling·14h

16/31 I think this is one of the things that made model customization much easier for individuals and small teams. Current cloud GPU examples: • RTX 4090: from around $0.69/hour • A100: from around $1.39/hour So a 10-hour training run could start somewhere near $7 to $14 in compute.

English

Nash A. | That Upgrade Feeling@UpgradeFeeling·15h

@unkonfined Eliminate negative people from your life because they hold you down.

English

Unkonfined@unkonfined·16h

It's important to remove anyone or anything from your life to protect your peace.

English

365

291

1.1K

29.5K

Nash A. | That Upgrade Feeling@UpgradeFeeling·15h

15/31 The fifth way is LoRA or QLoRA. Instead of updating billions of model weights, LoRA trains a much smaller adapter while most of the model stays frozen. QLoRA also compresses the model, often to 4-bit, which reduces the memory needed during training.

English

Nash A. | That Upgrade Feeling retweetledi

Apple Hub@theapplehub·21h

Apple will officially release iOS 27 next month for these supported iPhone models 🔥

English

457

5.5K

347.4K

Nash A. | That Upgrade Feeling@UpgradeFeeling·1d

14/31 A simple way to understand it: If the problem is, “The model does not know my documents,” then use RAG. If the model already knows the answer but does not follow your style or format, fine-tuning may help. And obviously, many strong systems combine RAG, fine-tuning, and tools together.

English

Keşfet

@NanouuSymeon @elonmusk @BarackObama @taylorswift13 @cristiano @BillGates @NASA @nikifrancismediavine