I Built the Only AI Resource You'll Ever Need, and It's Free (SuperBash Learn)

Learning AI is hard for a dumb reason: there is too much information, and most of it does not point you at tools you will actually use. So we spent the past few weeks building SuperBash Learn, a free site that pulls together everything we have found useful into one place. It is not a course. Nobody needs a CS20-style syllabus to get productive with AI. You need to know which tools exist and when to reach for them.

Start with an agent, not a textbook

The best starting point on the site is the Hermes guides. We have full walkthroughs for getting Hermes up and running, and the reason we push it first is simple: Hermes can run on a remote server 24 hours a day and do work for you while you sleep. Learning AI through an agent model teaches you the real workflow, not toy prompts in a chat box. Every one of our videos also has a full written guide on the site with copy-paste commands, so you can follow along and then refer back later without scrubbing through footage.

The tier list is opinion, and that is the point

We keep a model tier list on the site that we update as new models drop, which right now is basically every week. It is not sponsored, and no amount of money changes the rankings. But be clear about what it is: our opinion from daily use, not a numbers dump. If you want numbers, AI bench exists. What we rate is how a model performs day to day for real people doing real work.

Fable 5 sits at the top because it is the smartest model we use and gives genuinely good answers when you ask it hard questions. Right behind it in S tier we have Sol and Kimi K3. K3 earns its spot even though it is not always the smartest and can burn more tokens, because it is really, really cheap. Cost per task matters, so it goes into the ranking. That is the kind of judgment a benchmark table will never give you.

See models before you burn hours on them

The visual benchmarks section runs the same one-shot prompts, like our Helms deep plan visualization, across every model. You can see the output side by side and get a feel for what a model actually produces. Sometimes a model just makes something worse, and you can spot that in five seconds of looking instead of reading a leaderboard.

The compare models page is meant for the same job, though I will be honest: it needs work, and I am going to improve it. The goal is a side-by-side view so you can decide whether a new model deserves your time. That decision matters more than people admit. Every new release is exciting, but if it takes you two or three hours to get a grip on it, you want to know it is good before you invest the evening.

Skills are the sleeper feature

The last pieces are skills and tools. Our marketing guy, who is a genius at this, collected the prompts he finds most useful for marketing, brainstorming, and design, and we host them as a curated skill repository. We did not write these and we are not taking credit; we just found they are the ones that work.

A skill feels weird at first because it is technically just a document that tells the AI what role to play. In theory the model already knows this stuff. In practice, skills genuinely improve output, because having the skill activates that part of the model. Ask an AI to do SEO bare and it is noticeably weaker than the same model with the SEO skill loaded. One of my colleagues built an entire website, a very polished one aimed at Hong Kong and written in Chinese, using just these skills and prompts. We may do videos on getting the most out of this if people want them.

The site has no ads. At most, there are referrals for tools we already use ourselves. It is not a money enterprise, and my team is not 100 people. We built as much as we could, and we will keep adding to it. If you find it useful, use it. If you do not, leave a comment and tell us, because if it is not useful we should not bother.