18:31Alex Finn · Tutorial
How to get unlimited AI for free (GLM 5.2 local)
An 18-minute field report on running a frontier-class open-weights model locally -- for free, forever, with zero cloud costs.
June 19thThe argument in one line.
When frontier AI access is rationed by governments and executives rather than earned by builders, the technology stops being a meritocracy and starts being a permission system — and that shift demands a local sovereignty strategy.
Read if. Skip if.
READ IF YOU ARE…
- You rely on frontier AI models for business work and felt the sting when Claude Fable 5 was pulled from public access.
- You're considering investing in local AI hardware but don't know where to start or what hardware tier fits your use cases.
- You follow AI news closely and want a practitioner's take on what ChatGPT 5.6's gated rollout means for everyday builders.
- You're already running local models (Ollama, LM Studio, etc.) and want to understand the memory-vs-bandwidth tradeoff between Mac Studio, DGX Spark, and RTX 5090-class builds.
SKIP IF…
- You want a technical deep-dive into ChatGPT 5.6's architecture or benchmark methodology — this is a reaction take, not a technical review.
- You're not interested in building a home lab and are content waiting for broad cloud access.
The full version, fast.
ChatGPT 5.6 arrives in three tiers (Sol, Terra, Luna), outperforms Claude Mythos at one-third the cost, but is locked to 20 companies chosen by OpenAI and the US government. The presenter argues this is the clearest sign yet that frontier AI is becoming gated infrastructure — not a free market. His three-step response: keep using ChatGPT 5.5 with Codex as the agent harness (still the best available), start investigating local AI hardware seriously, and build a home lab once you know your use cases. He demos his own setup (Mac Studios, DGX Spark, RTX 5090) and breaks down the memory-vs-bandwidth tradeoff across three hardware tiers so viewers can choose intelligently.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:05
01 · Cold open — "worst news of the year"
Hook delivered directly to camera. ChatGPT 5.6 beats Mythos, costs 1/3 as much, available to 20 companies. The title provocation explained.
01:05 – 01:53
02 · Sol / Terra / Luna breakdown
Slide showing three model tiers. Sol = flagship agentic; Terra = thinking mid-tier; Luna = smaller fast model. None available to the public.
01:53 – 03:53
03 · The permanent underclass argument
America decelerating innovation for the first time. Government review gates access. Boxing ring metaphor: one boxer with bricks, the other hands-tied. Winners are handpicked, not earned.
03:53 – 04:43
04 · Polymarket thread of hope
OpenAI says broad access in coming weeks. Polymarket shows 85–89% odds of Claude Fable 5 returning by July 17. July is the target month.
04:43 – 06:40
05 · WHAT TO DO — the three-step prescription
Slide: (1) ChatGPT 5.5 still best right now — use it via Codex; (2) Research local AI; (3) Build a home lab if comfortable. Soft CTA for engagement.
06:40 – 07:40
06 · Home lab tour
Shows custom local AI control plane dashboard with Jobs, Running Models, and Findings panels. Lists hardware: 3x Mac Studio 512GB, 2x Mac Mini, DGX Spark, RTX 5090 (64GB RAM). Agents run 24/7.
07:40 – 09:32
07 · LOCAL MODELS comparison — three hardware tiers
Slide: Mac Studio (high memory, low bandwidth), AI Computers/DGX Spark (128GB medium, medium bandwidth), Powerhouse chips/5090/6000 Pro (low VRAM, very high bandwidth). Pick based on your use case.
09:32 – 10:32
08 · Reverse prompting + outro
Strategy: before spending money, ask your AI agent what hardware fits your use cases. Closes with acknowledgment that all access paths are constricting, then hard CTA.
Lines worth screenshotting.
- ChatGPT 5.6 is better than Claude Mythos and costs one-third the price — and you cannot use it.
- Restricting the best AI to 20 hand-picked companies converts a meritocracy into a permission system where winners are chosen, not earned.
- Frontier intelligence is the most valuable resource ever created; whoever controls access to it controls who gets to win.
- A boxing match where one fighter has bricks tied on their hands and the other has arms pinned behind their back is a fair description of gated AI access.
- Codex as an agent harness makes ChatGPT 5.5 more effective than Opus used raw — the harness can matter more than the raw model.
- Before spending $10,000 on local AI hardware, reverse-prompt your existing AI agent to identify your actual use cases — then buy.
- Mac Studio wins on memory capacity (up to 512GB unified), loses on bandwidth; RTX 5090 wins on speed, loses on VRAM (32GB); DGX Spark hits the middle at 128GB with medium bandwidth.
- GLM 5.2 at 250GB weights requires a Mac Studio — NVIDIA cards simply don't have enough VRAM to load it.
- The Polymarket prediction market put 85–89% odds on Claude Fable 5 returning by July 17, 2026.
- Building a local home lab is a sovereignty move: governments and executives cannot gatekeep a model running on your own hardware.
- As hardware prices rise, the local sovereignty path gets harder — the window to build cheaply is now, not later.
- Broad access to ChatGPT 5.6 is expected within weeks; the temporary access gap favors the 20 selected companies in the interim.
How to stay competitive when frontier AI gets gated.
WHAT TO LEARN
Restricted access to the best AI models is not a temporary glitch — it is a structural shift, and the response is to diversify toward tools and hardware you control.
- When a new model ships behind a permission wall, the best available public model remains more useful than no frontier model at all — today that means ChatGPT 5.5 via an agent harness like Codex.
- Local AI hardware is a sovereignty hedge, not a replacement for cloud AI — the goal is having access the market cannot revoke, not running the fastest possible model.
- Mac Studio wins on memory capacity (run massive models), DGX Spark wins on plug-and-play setup (medium models, medium speed), and discrete GPU cards (RTX 5090, 6000 Pro) win on inference speed for smaller models.
- Before purchasing any local AI hardware, determine your use cases first — ask your existing AI agent what it would recommend for you based on what it already knows.
- Polymarket and similar prediction markets are underused signals for timing tech decisions; the crowd was pricing Claude Fable 5 restoration at 85–89% odds by July 17, 2026.
- Reverse prompting — using your AI agent's existing context about you to get personalized recommendations — is more useful than generic research when the decision space is wide and expensive.
- The window to build local AI sovereignty cheaply narrows as hardware prices rise and regulation tightens; the time to investigate is before you urgently need it, not after.
Terms worth knowing.
- Sol / Terra / Luna
- The three tiers of ChatGPT 5.6. Sol is the flagship agentic model (equivalent in scope to Claude Fable 5). Terra is the thinking-mode mid-tier that reportedly beats ChatGPT 5.5 at a fraction of the cost. Luna is the smaller, faster variant for high-volume tasks.
- Frontier intelligence
- The presenter's term for the most capable AI models available at any given moment. Used to distinguish top-tier models from commodity or open-weight alternatives.
- Permanent underclass
- The presenter's recurring framework: when access to transformative resources (here, frontier AI) is controlled by a small group, everyone outside that group structurally cannot compete — and the gap compounds over time.
- DGX Spark
- NVIDIA's plug-and-play AI computer with 128GB unified memory. Positioned between Mac Studio (higher memory, slower) and discrete GPU builds (faster, less memory). Currently popular because of its turnkey setup.
- Hermes agent
- A personal AI agent the presenter has built with deep context about himself and his business. He recommends asking it hardware questions before making purchasing decisions.
- Reverse prompting
- Strategy of asking your AI agent what it recommends for you, based on everything it already knows about you, rather than starting from a blank search. Useful before large purchases like local AI hardware.
- Unified memory
- Architecture used in Apple Silicon and some AI computers where CPU and GPU share the same memory pool, enabling larger models to be loaded than discrete VRAM cards allow.
- VRAM
- Video RAM — the on-chip memory on discrete GPU cards. Limits the maximum model size that can be loaded; RTX 5090 has 32GB, which constrains it to smaller quantized models despite very fast inference speed.
- Polymarket
- A prediction market platform where users bet real money on outcomes. The presenter uses it as a signal for when Claude Fable 5 will return to public access (85–89% odds by July 17).
Things they pointed at.
05:20toolCodex
06:00productMicro Center
07:54productDGX Spark
09:45toolHermes agent
Lines you could clip.
01:03
“There's a small group of winners, and you ain't one of them.”
02:59
“It's like getting into a boxing ring where one boxer has bricks tied on their hands and the other has their hands tied behind their back. The one with the bricks is gonna dominate.”
02:59
“Frontier intelligence, because when you use this, you can build and do anything you want.”
06:13
“I wanna invest in a home AI lab where governments and executives cannot take away my AI, where I can build my own sovereign intelligence.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogystory
ChadGPT 5.6 just got announced, and it is legitimately the worst news of the entire year.
Let me explain. And, no, it's not because it was named after three failed cryptocurrencies. No.
That's not it at all. Few things to note here is, one, it is better than Mythos. So looking at the benchmarks, it beats Mythos on terminal bench.
On top of that, it's also one third of the price of Mythos. So it's better than the best AI model you've ever used in your life. It is a third of the price, but here's where it absolutely blows.
It's not available to you. You can't use it. It is only available to 20 companies in the world that OpenAI and the US government has hand selected to win.
That's right. Much like Fable five, you cannot use the best technology out there anymore. The people that determine who can use the most powerful technology ever created are the CEOs and the government.
That's it. There's a small group of winners, and you ain't one of them. So, yes, there will be three versions of chat GBT 5.6.
Soul, which is the big mamma jama, the fable five version. Terra, which is like their down the middle five five thinking version, which apparently beats five five for a fraction of the price. And there's Luna, which is their smaller model, which comparably is their smaller model, but it's not nearly as small as, like, their micro or mini model.
So I guess the micro models are getting bigger now. Those are the three versions. They're all great, and we won't be using it anytime soon.
In a second, I'm going to tell you what you need to do about this, what your next steps are to take advantage of this moment. But let's talk about why this is absolutely horrible news for America and absolutely horrible news for the world. America for the first time is decelerating their innovation, decelerating their technology.
Unlike many countries out there, there weren't many rules or regulations. You can build new technology, put it out there, make money off it, and innovate quickly. This allowed us to be the number one country in the world when it came to AI models.
We had the best AI models on planet Earth. Now for the first time, we're being decelerated. You have to review the models with the government.
And if the government decides it's not safe in your hands, you don't get to use it. Now this isn't a political channel.
I'm not getting into politics, but I think taking politics out of it, you can objectively say this is not good. When governments and executives can decide who winners are, who can use the great technology, who can use the crap technology, that is not good for anyone.
Winners shouldn't be handpicked. It should be a meritocracy. Everyone should have equal access and opportunity, and the ones that use it the best are the winners, and they earn their right to win.
That's no longer the case. I have been talking about the permanent underclass for a long time now. This has pissed off many, many people, but this is what the permanent underclass is.
When certain people get access to certain resources, and this is the most valuable resource ever created, , because when you use this, you can build and do anything you want.
Those people end up being winners, and everyone else ends up being losers. If you are building a business and only you have access to the best intelligence and everyone else has no access, you're going to dominate them.
They will have zero shot. It's like getting into a boxing ring where one boxer has bricks tied on their hands and the other has their hands tied behind their back. The one with the bricks is gonna dominate.
Up until this point, this was all democratized. Anyone can use this amazing technology. Now it's not democratized.
Now it's handpicked winners. So it's not about who's the best, who's working the hardest, who has the most skills.
No. It's not about who's contributing the most to society. It's about who's the best friends with the executives and the government.
That absolutely sucks. I was hoping with Fable five being taken away, maybe Chad GBT can somehow sneak their way to being the good guys and be able to release our models.
And listen. I get it. This isn't up to them.
This isn't up to OpenAI. It wasn't their choice to hold this back. After seeing what happened with Anthropic, I'm sure they was like, no.
We don't we don't want any part of that. We don't wanna do that. We'll just hold it back and be safe.
But this absolutely sucks. We are all losers. OpenAI did say this will be available broadly in the coming weeks, so I am hoping by the end of July, we're able to get access to this.
And if you look at the poly market, it's looking like by July 17, we will get Fable five back. That is very good news. Apparently, people with a lot of money I'll get this crap off my screen.
People with a lot of money are betting on by end of July getting Fable back. So it looks like July is the month we get our hands on the frontier technology. The challenge is there's gonna be a lot of people, a lot of businesses who are hand selected, who will be using this until then and will have a distinct advantage on us.
What does that mean for you? What should you be doing? Well, let's go to this slide over here, I'm gonna go into more detail.
So stick around because by the way, if learned anything so far, leave a like down below. Subscribe. Turn on notifications.
All I do is create amazing content about AI and keep you up to date with what's going on. Uh, number one, I still think five five is the best right now. I think five five is the best model.
I believe that because I think Codex is the best agent harness. I actually think Opus might be a better model, but Codex is so good at controlling your computer, doing computer use, browser use, all that. I believe it makes five five better.
So five five is the model I'd be using right now. Here's where things switch up a little bit. I'd research local AI.
What do I mean by that? I just went out to Micro Center. Hey, Micro Center.
Sponsor me. I went there. I built this computer you see here.
It's sitting down there. I'm be playing Cyberpunk on it tonight. It is an RTX fifty ninety build.
It's got 64 gigs of RAM. It would have cost me, like, 2,000 more dollars to make it one twenty eight gigs of RAM. Anyway, I built this computer.
I am building a home lab. If you've watched this channel, if you follow my Twitter, you know I've already been building a home lab. I have three Mac Studio five twelve gigabytes, two Mac Minis, and a DGX Spark.
Now I have this RTX fifty ninety. I am building my own home lab so I can run the most powerful models on planet Earth. I want to invest in local models.
I wanna invest in a home AI lab where governments and executives cannot take away my AI, where I can build my own sovereign intelligence, build my own systems, and no one can control it. No one can gatekeep it.
Nothing like that. I believe everyone should be investigating local AI right now. Do I believe you should go out and spend $10,000 on a computer right this moment?
No. I do not believe that. I think that is a mistake.
I think what you need to do first is research it, figure out what's suitable for you, what models are suitable for you, what hardware is suitable for you, most importantly, what you would do with that AI, which use cases you would do. I'm gonna go into all that in a second on how to figure that out. Before you spend the tens of thousands of dollars, first, figure out what you're gonna do with it.
So when you set it up, you can start getting cooking. I'm gonna show you what I'm doing in a second. Actually, let's just get straight into it.
This is my local AI home lab platform I am building at the moment. This allows me to monitor all my devices. It also allows me to see which models I have up and running.
And these models are going twenty four seven, three sixty five. They're doing security scans. They're checking code.
They are writing code at all times. They are constantly doing things around the clock for me, pushing my business forward, pushing Henry Intelligent Machines forward. And I am building on top of this home lab so I can do more and more things.
I plan on training models, training lures, many different cool things like that. You want more information on that? Let me know down below.
Would you want me to, like, walk through how I'm building this home AI lab, the hardware I'm using, what I'm doing with it? Let me know down below in the comments. So if you're curious what you should be doing when it comes to local AI models, should I be doing Mac Studios, Mac Mini?
Should I be buying a DGX Spark? I will spend a couple minutes here. So lock in for this.
You basically have three options. You have high end Macs. You have AI computers like DGX Sparks and the new AMD computer coming out as well as powerhouse chips like the fifty ninety I just built, like the 6,000 pro I'll just be adding on to it soon enough.
What's the difference in each? Which one should you go for? Well, couple things here.
The advantage to max are high unified memory. That means you get ton of memory to load models into. So you can run the biggest models.
I'm running g l m five two on one of my Mac studios right now. It's 250 gigs in my memory, and I can only do that on my Mac because NVIDIA chips don't have that much vram. So Mac Studio's great for massive big models.
It's just low memory bandwidth, which means it's very slow. AI computers like DGX sparks, which are very popular right now because they're plug and play. You go to the store, buy, you plug it in, you're good to go.
Those have medium unified memory, 128 gigs, so you can load medium sized models like really good QUEN three six models. And they also have medium bandwidth, so they're much faster than Mac Studio. Not lightning fast, but faster.
Then you have the Powerhouse chips, the 50 nineties, the 6,000 pros. These are lower VRAM. The fifty ninety can only store 32 gigs.
So kind of the smaller quen models. The 6 thousand's much bigger. It's like a $12,000 chip.
But their advantage is they have very high bandwidth, so you can run models lightning fast on them. I bought a computer in each class so I can take advantage of all the strengths and weaknesses. So it's up to you.
If you're getting into local models, think about which one you want. Think about what suits your use cases. If you're not sure what suits your use cases, go to chatgbt.com.
Go to claud.com. Go back and forth with your agent. Go to your Hermes agent.
Say, hey. Based on what you know about me, what would you do? What use cases can we do?
What hardware should we get? What should we do with these local models? Which one should we load?
Your agent should be able to help you out because it has a ton of context around you. Make sure to be reverse prompting as much as you can before you go out and spend tens of thousands of dollars on computers. This is the news of the moment.
It quite honestly pisses me off, but I think brighter days ahead. I hate that this is the new normal. I hate that this is world we live in at the moment where you no longer have access to frontier intelligence.
You need to be getting permission first. I'm sure this will only continue.
And as hardware prices get more and more expensive, it's gonna be harder to become sovereign, run things on your own. So all options are starting to become more constricted, unfortunately.
This is what's going on. More videos coming in the next couple days on Hermes use cases, Claude code best practices, things like that. So make sure to lock it in, leave a like, subscribe, turn on notifications, hit all the buttons you see on your screen.
So grateful you'd watch. Hope this was helpful. See you in the next
The Hook
The bait, then the rug-pull.
A model that beats the best AI you've ever used and costs one-third as much was announced this week — and you cannot have it. Alex Finn opens with barely-contained fury before laying out the logic: ChatGPT 5.6 isn't bad news because of what it is, but because of who gets to use it.
Named ideas worth stealing.
07:40list
Local Hardware Tier Matrix
- Mac Studio — high unified memory (up to 512GB), low bandwidth; best for massive models like GLM 5.2
- AI Computers (DGX Spark, AMD Halos) — 128GB unified, medium bandwidth; plug-and-play, good for Qwen 3.6-class
- Powerhouse chips (RTX 5090, 6000 Pro) — 32GB–large VRAM, very high bandwidth; fast inference on smaller models
Three-tier hardware framework for choosing local AI equipment based on the memory-vs-bandwidth tradeoff and your model size needs.
Steal forAny conversation about local AI setup — helps someone pick hardware before spending $10K+
09:45concept
Reverse Prompting for Hardware Decisions
Before purchasing local AI hardware, ask your existing AI agent: 'Based on what you know about me, what use cases should I run locally and what hardware fits?' Leverages the agent's existing context about you.
Steal forAny expensive tech purchasing decision where the buyer has an established AI agent with personal context
02:40concept
Permanent Underclass Framework
When access to the most powerful resource of an era is gated by a small group, those outside structurally cannot compete — not because they lack skill or effort, but because they lack permission.
Steal forAny content about AI access inequality, government regulation of tech, or sovereignty arguments
How they asked for the click.
VERBAL ASK
09:45subscribe
“leave a like, subscribe, turn on notifications, hit all the buttons you see on your screen”
Standard verbal CTA repeated twice (mid-video at ~4:55 and outro). No on-screen graphic. Mid-video ask is embedded inside the action section for natural flow.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
