Crof AI

Here on X we live in a bubble, people out there have no idea of what Hermes Agent is, or Local AI, or llamaccp. They just know ChatGPT, Claude or Gemini. Unbelievable to me, what we give for granted, it's totally unknown for the vast majority out there!

Deepseek v4 pro is 27x cheaper than GPT5.4 and 31x cheaper than Opus 4.6. BUT it's not perfect. It has flaws. How to get Maximum from DeepSeek v4? 1> Use Codex 5.5 as brain and Deepseek v4 as muscle. Codex plans the work, DeepSeek produces the work, and Codex reviews and finalizes the work. 2> Write specs: start every project with a Codex plan Using Deepseek v4 without specs is a crime! Codex should define the job clearly. Write PRDs, attach schemas, rule files, provide strict output formats. 3> Force DeepSeek to stay grounded DeepSeek’s biggest weakness is that it can invent details that sound believable (classic hallucination). To prevent this, always tell it that every factual claim must come from the provided source. 4> Give DeepSeek a clear output box DeepSeek performs much better when the output format is strict. Give it a JSON schema, a table format, exact field names, examples of good output, examples of bad output, and a checklist it must pass. 5> Use only for hard problems When you provide specs with strict workflow to follow you don't need reasoning mode, that slows it down. The rule is: use cheap mode for volume and thinking mode for difficulty. 6> Batch similar tasks together DeepSeek works best when one batch contains one kind of task. With batching it's cheaper and more organized but make sure you have quality guardrails to evaluate the outputs. 7> Use Codex to build quality checks For repeatable work, Codex should create validators. These checks should confirm that JSON parses correctly, fields are not empty, placeholders are removed, duplicates are detected, etc. 8> Never let DeepSeek guess critical values DeepSeek can invent emails, phone numbers, file paths, user IDs, database IDs, event IDs, API names, version numbers, prices, dates, citations, etc (hallucinations). 9> For long context, ask for an index first When giving DeepSeek a huge document, repo, or log file, do not immediately ask for the final answer. First ask it to create a map of the material: key sections, important details, source locations etc. 10> Use its weakness for creativity DeepSeek’s tendency to create plausible details is dangerous for factual work, but useful for creative work. Use it for brainstorming, naming, scenarios, UI copy options, test ideas, edge cases, and synthetic examples. > Final Principle DeepSeek V4 gives the most value when you treat it as a powerful worker, not an unquestioned authority. - Let DeepSeek generate aggressively. - Make Codex verify carefully. - Keep what passes the validator. - Repair or discard the rest.

I actually think the whole "permanent underclass" narrative is wrong. I think we're about to see the largest EXPLOSION of entrepreneurship in human history. I get why the fear exists. Jobs are getting cut. AI researchers are privately saying most people are screwed. The models are getting ridiculously better and faster than anyone expected. Project that forward linearly and yeah, it looks BLEAK. But linear projections are usually wrong during platform shifts. Nobody projected that the internet would create 50 million small businesses. They projected Walmart would eat everything. Nobody projected that mobile would create a million app developers. They projected phones were just phones. What actually happens is intelligence gets cheap and a flood of new builders enter the market with domain knowledge the incumbents never had. Millions will get laid off or just never hired over the next 24-36 months. Those jobs are not coming back. So they become entrepreneurs. Out of necessity at first. Then out of opportunity. The underclass idea is VIRAL because it confirms something people have been feeling for a decade. That the ground is shifting and nobody at the top is reaching down. And they're right. But the interesting thing about this particular technology is that it doesn't check your resume or your zip code. The same tool that eliminates your position hands you the ability to build the thing that replaces it. The weapon and the escape hatch are the same object. We're about to see more new companies started in the next 5 years than in the previous 50. And I think we're going to look back at this moment the way we look back at 1995. Everyone was scared. Everyone was right to be. And the people who built anyway became the next generation of owners. I know you might be reading about the permanent underclass and it's scary. Who wants to "get stuck in the permanent underclass no one. My POV is the permanent underclass isn't a foregone conclusion. I know some people are genuinely struggling right now and "just go build" sounds tone deaf when you're worried about rent. I get that. But the reason I'm optimistic is that the cost to start something just dropped to nearly zero, intelligence on tap, and eveyr category/industry you can think of is getting reshuffled. The explosion of entrepreneurship is just beginning.

if you wanna get a $20 monthly plan for vibe coding definitely don't get Claude, not even Codex. you should get: - GLM 5.1 - Minimax 2.7 - Kimi 2.6 - or Deepseek v4 All these will give you so much more usage for the buck, and you won't get stopped at the start or mid-session.