Getting started in AI coding.

Many of us started coding long before AI. You had a problem, you dug in, and solving it was half the fun. When you got stuck, you went to your sources:

The Python Standard Library docs → https://docs.python.org/3/library/index.html

Later, Stack Overflow, where others posted snippets you could copy, paste, and adapt. Almost always helpful and you just say thanks. It just worked.

Then came AI. Suddenly, instead of hunting through forums, you could ask a model directly and get a snippet back. Copy, paste, done. Even better paste your full code, AI could often debug your code, point out problems you didn’t even know were there, and suggest fixes.

But copy-paste has limits. Sometimes your code snippets are missing key context. That’s where agentic coding comes in. Instead of copy-pasting back and forth, you let the AI read the files that matter, handle the context, and even write back into your codebase.

Subscriptions

The best solutions today use state-of-the-art (SOTA) models from large providers. They’re powerful, but they come with tradeoffs: Monthly fees with vague usage limits; No real privacy.

Some popular setups include:

Codex CLI + GPT-5 Codex

Gemini CLI + Gemini 2.5 Pro

Claude Code + Claude Sonnet 4.5 (arguably SOTA #1 right now)

Cursor

Pricing is usually $25/month (reasonable limits) or $250/month (much higher limits). If you’re a developer running AI for hours a day, those limits vanish fast. Running models locally avoids these restrictions and keeps your data private.

Budget vs Effectiveness

Let’s compare hardware vs subscriptions:

$600 (16GB VRAM GPU) → Equal to ~2 years at $25/month, or ~5 months at $250/month.

$2,500 workstation → ~8 years at $25/month, ~1 year at $250/month.

Of course, cloud models outperform local ones:

GPT-5 high → 85% (LiveCodeBench)

Grok-4 fast → 82%

Gemini 2.5 Pro → 80%

$10,000 local rig (large model) → ~79%

$3,000 local rig (GPT-120B) → ~65%

$600 local rig (GPT-20B, 16GB GPU) → ~57%

Here’s the catch: 57% is still excellent. That’s roughly equal to a developer with ~5 years of experience. A top-tier developer might match SOTA models, but most developers won’t.

Subscriptions can’t play video games or run TensorFlow, PyTorch, or ComfyUI. Hardware can.

Hardware Options

Option 1 — Upgrade an Existing Gaming Rig

GPU is everything. 8GB GPUs are fine for gaming but limited for AI. You can run small models (e.g. Qwen3-4B), but that’s not ideal for beginners.Upgrade to 16gb of vram.

If you already have a 16GB GPU, you’re set. Dont fall into the hardware trap of MORE VRAM.

GPT-20B (FP4, 80k context, flash attention) runs great on 16GB.

~$600 for AMD 9060XT or Nvidia 5060 (16GB VRAM)

Advanced option for later: dual GPUs for 32GB VRAM + 64GB+ system RAM → allows 10–15 TPS on large MoE models.

Option 2 — No Existing Setup (~$2,500)

If you’re starting from scratch or your workstation is ancient and needs full replacing.

Mac Mini M4 Pro with 64GB unified RAM

AMD Strix Halo (Ryzen 395 AI Max+) — ~$2,400 for 128GB unified RAM

Performance expectations for either machine.

11 TPS → Qwen3-235B

20 TPS → GPT-120B

55 TPS → Qwen3-30B

14 TPS → LLaMA-4 109B

Future Hardware

As mentioned above, don’t go chasing more than 16GB right away. There’s some upcoming solutions that will be ideal. Hold your horses and wait for these if you can. Solid option to pay for a subscription now and wait for these.

AMD Medusa Halo (ETA ~2027) → DDR6, higher bandwidth, more cores, hopefully 256GB unified RAM.

Nvidia DGX Spark → 128GB unified RAM, ~1 petaflop @ Q4 precision. Rumored ~$6,000 coming out this month?

Huawei Ascend 310 AI Studio Pro → 192GB RAM, competitive with Strix Halo/Mac Mini.

Huawei Atlas 300I Duo → 96GB VRAM for ~$2,000, but slower memory bandwidth.

Software

LM Studio (closed source, but polished GUI): https://lmstudio.ai/

Once you’re set up, you’ll download models often. For 16GB VRAM, start with GPT-20B.

npm install -g @openai/codex

then in in ~/.codex/config.toml

[model_providers.lmstudio]

name = "LM Studio"

base_url = "http://localhost:1234/v1"

[profiles.gpt-oss-20b]

model_provider = "lmstudio"

model = "gpt-oss:20b"

On the Cli:
codex --profile gpt-oss-20b

You’re now cooking.