
Your AI Stack Isn't Ready for Claude Mythos
What this covers
This is a single-speaker commentary, organized around the leak of Claude Mythos, Anthropic's next model. The speaker's claims about it are that it was trained first on Nvidia's GB300 chips, that it is a step change rather than an incremental gain, and that release is likely within a month or two. The security evidence comes first: researchers testing it ahead of release reportedly found previously unknown vulnerabilities in Ghost, a 50,000-star GitHub repo, that skilled human researchers had missed. The rest of the video turns that into four audit questions for anyone building on weaker models, covering prompt scaffolding, retrieval, hard-coded domain knowledge, and verification.
Each question gets its own argument. On scaffolding, the speaker suggests asking of every line in a prompt whether the model needs it or whether a human needed the model to need it, and moving toward outcome specification over process steps. On retrieval, the claim is that responsibility should shift from hand-designed RAG toward letting the model choose what to pull from well-organized repositories. The section on domain knowledge argues that prompting skill is becoming a matter of what you leave out. The verification section says end-of-pipeline evaluation must cover functional and non-functional requirements, edge cases, and code quality, and that human handoffs in an agentic pipeline will become the bottleneck. Constraints that survive upgrades, such as never disclosing customer financial data, are set apart from the cruft that doesn't.
The later sections are more personal and contentious. The speaker argues that process is bound up with work identity and has to be let go, and that a model given a business process, whether Mythos or Andrej Karpathy's auto research, will find optimizations humans can't. He also treats premium plan access as buying a "better brain," says model capability may matter more than individual talent, and forecasts non-technical people building their own team and family software through chat.
Claude Mythos represents a critical inflection point requiring immediate simplification of AI systems—removing unnecessary scaffolding, process specifications, and hardcoded knowledge so more capable models can operate with greater autonomy and efficiency.
- Larger models force simplification because they can infer context and handle complexity without explicit instructions, making detailed prompts and procedural steps counterproductive
- The 'bitter lesson' of LLM development is that simpler systems outperform complex ones as model capability increases, requiring a fundamental shift from specifying 'how' to specifying only 'what' and 'why'
- Organizations that fail to simplify systems and shift from compensating for model limitations to architecting intelligent workflows will lose competitive advantage as next-generation models arrive within weeks
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Claude Mythos is the first model trained on Nvidia's new GB300 chips, represents a step change in capability rather than incremental improvement, and is likely to be released within the next month or two.
“Claude Mythos is the first model, as far as we know, that has been trained on Nvidia's new GB3 chips. It is a massive model. It is a step forward. Anthropic has confirmed its existence, and they've given it a new lineage name. So, it won't be called Sonnet, it won't be called Opus. It appears to be called Capybara.”
Anthropic is taking the unusual step of allowing security researchers to battle-test Claude Mythos against popular software utilities before release to harden defenses in advance.
“Anthropic is taking the unusual step of allowing security researchers to dig into Mythos, to battle test Mythos against some of the most popular utilities on the internet, and to harden up their defenses ahead of time”
When models get more capable, they force simplification of systems because the model can now accomplish much more than before, making prior workarounds and scaffolding unnecessary.
“when models get bigger, they force you to simplify. They force you to think, what can I delete about my systems and my practices because the model can do so much more now that that it couldn't do before.”
Simplification is hard because most people have added guardrails and cruft around work to feel valuable or to compensate for weaker models; the challenge is to grow by simplifying what you ask the model to do.
“And yes, you absolutely will need to look in detail to get there because most of us have added a lot of guardrails and a lot of cruft around our work so that we feel like we add value and or so that the weaker models can do their jobs. That's fine. The challenge is to grow by being able to simplify what you ask the model to do against a larger outcome set so that we get out of the way and let the model do its job, and we focus on [30:03] making sure that we are aiming the model in the direction of a big cool goal,”
Constraints and guardrails should be business rules that remain valid across model upgrades (e.g., 'never disclose customer financial data', 'always verify refund eligibility against policy') and should survive the shift to outcome specifications.
“Number two, think about your constraints and guardrails. These are things that must be true regardless of how the model achieves the outcome because when you give the model more flexibility, you need to be more clear about your constraints and guardrails. For example, never disclose customer financial data. That's a pretty good customer guardrail. Always verify refund eligibility against our policy. There's another one. These should survive model upgrades because they represent ongoing business rules that you want any model to follow regardless of how smart it gets.”
For non-technical users working with LLMs like Claude, effectiveness improves by stating desired outcomes in plain language and explaining why, rather than specifying how the model should accomplish the task.
“Ask for what you want in the end, and explain why. In plain language, and you don't need to elaborate on how to get there. As long as the model has access to the inputs it needs, the data it needs, it's going to get there on its own.”
With more capable models and larger context windows, humans should move retrieval logic responsibility from human-specified retrieval augmented generation (RAG) systems toward allowing the model to intelligently select what to retrieve from well-organized repositories.
“In the past, we had to carry a lot of the logic for retrieval on our side, so the model would retrieve correctly. But really, how much of my retrieval logic should belong to the model if the model's smarter?”
Human talent alone will not overcome productivity disadvantage from lacking premium AI access—increasingly, model capability matters more than human talent in determining work quality and speed.
“You might think, 'Oh, but talent will make up for it.' No, human talent will not. Like increasingly, the whole point of human talent is to simplify and get out of the way so that AI can do its thing. And the bigger the models get, the more that becomes obvious.”
Individuals and teams should assess their current roles to determine what portion is 'compensating for model limitations today' versus 'architecting and aiming AI correctly', with the goal of transitioning work toward the latter.
“you should be stepping back and asking yourself, 'How much of my role is compensating for a model's limitations today versus how much of my role is really thinking about how to architect and correctly aim artificial intelligence so it accomplishes a lot of work?' That's the part you want to be on, right?”
The way humans have described process as precious and tied to work identity—the specific series of steps we execute—is something we must let go of as more capable models become available, because what matters increasingly is only the outcome, not the procedural steps taken to achieve it.
“All of the way we have described process, the things that are precious to us, are things that are associated with our ability to execute work in a certain series of steps, and somehow we've decided that's an important reflection of our work identity. What Claude Methos and similar models are going to teach us is that that doesn't matter anymore. And what matters is the outcome, and our ability to name the outcome, and let go of the process.”
The art of prompting is evolving from 'what you put in' to 'what you leave out'—the skill remains valuable but is being redefined by model improvement.
“increasingly the art of prompting is about what you leave out. And it's still an art. It's still hard. You still have to prompt. And the skill is valuable. It's just that the skill is evolving because the models are getting better.”
If you give a business process to Mythos or Andrej Karpathy's auto research, you're likely to find ways to optimize it that a human cannot figure out, because LLMs are now better than us at finding efficient process execution.
“I would bet you if you give it to Methos, or maybe you give it to Andrej Karpathy's auto research, you're going to find ways to make that better that a human can't figure out. The LLMs are now better than us at finding efficient ways to process, and we should stop specifying.”
Non-technical knowledge workers will increasingly develop sophisticated 'under-the-desk' software applications using advanced models, creating team tools without formal engineering involvement.
“our under-the-desk software, our software we build as non-technical people is going to get increasingly sophisticated because we're going to have access to models that can do increasingly sophisticated things just by communicating an intent in plain chat.”
If your agentic software development pipeline depends on human handoffs or human review as a key step, you will be in trouble when models like Mythos arrive.
“If you are depending on humans and human handoffs as a key part of your agentic software development pipeline, you're in trouble. And there is sort of an analogy there for non-tech folks. If you are depending on humans as a handoff between PowerPoint and Excel and you're doing AI [17:10] for both of those pieces, you should be looking seriously at whether you can automate those non-tech artifact handoffs because Mythos can help you do that.”
Career planning should involve deciding whether to pursue the cutting-edge curve (pay for premium plans, be first, 10x leverage) or stay on a pro plan one step behind, accepting slower productivity and diminished competitive position.
“Think about it as what trajectory, what curve do I want to be on? Do I want to pay to be on the cutting edge curve? And I'm going to use it to the health, and I'm going to build really cool things, and I'm going to learn, and I'm going to leverage it to 10x myself, etc. Or do I [20:20] want to be a step behind on like a pro plan or whatever the 20 buck a month plan is? And then in that case, I'm willing to wait, and I'm not going to be the first. I'm willing to take the hit on my career for that.”
The pipeline supporting better model outcomes requires: clear goal specification, good data/tool availability, pipeline/support infrastructure, and high standards for output quality (clean code, accurate Excel).
“we focus on making sure that we are aiming the model in the direction of a big cool goal, and we focus on making sure we're building the pipeline and the support and the and the tool availability, and the availability of data so the model can do its job. And we also have a good sense of what good looks like. I've I've talked about that a lot. That's going to matter for technical and non-technical work. Is the code clean? Is the Excel absolutely right?”
Claude Mythos is exceptionally effective at finding zero-day vulnerabilities in software infrastructure, demonstrated by discovering undiscovered issues in Ghost (a 50,000-star GitHub repo) that even world-class security researchers had missed.
“one of the most experienced security researchers in the world stood up at a conference in San Francisco in the past few days and said that Claude Mythos immediately found zero-day vulnerabilities in Ghost, which is a 50,000-star GitHub repo that has never had major issues before. And as soon as Mythos was let loose on it, it found a bunch of issues that even the world's best security researchers haven't found.”
Families could increasingly build non-code family applications (like a shared family calendar) that integrate with services like Google Calendar, deploy automatically, and require no code editing because they're specified through chat with models.
“Let's say you want to build a family calendar. Like it's increasingly going to be possible to not just build one for your personal use that you start up personally on your laptop, but like that you just build and you can deploy and maintain for the family and it's not even going to be something you think about as a piece of software because you never touched the code.”
Prompt scaffolding should be audited at a per-line level, asking for each instruction whether it exists because the model needs it or because the human needed the model to need it, with the goal of removing unnecessary complexity.
“Number one is your prompt scaffolding, how you think about prompting to drive results. Ask yourself, and this is not a per prompt document question, this is a per line question. Really dig in. Is this instruction here because the model needs it or is it here because I needed the model to need it?”
The fourth preparation area is ensuring systems can verify that models have completed assigned tasks correctly through robust evaluation gates.
“question number four, this one's all about did the model do what we asked it to do? Can we sniff check it and show that it worked?”
Individuals and organizations should view premium AI plan investment as purchasing access to a 'better brain' and consider whether the productivity multiplier justifies the subscription cost.
“think about the kind of intelligence you want to purchase and the timeline you want to buy it on. Because the people who are able to get Claude Mythos now, if you have a plan to leverage it effectively, are going to be ahead. There's just no other way to put it. They have access to a better brain.”
Multi-agent coordination patterns where a primary planning agent (like Mythos) spawns specialized sub-agents is more effective than agent swarms, and Mythos-ready systems should implement hierarchical agent coordination.
“if you're thinking about larger agentic software, we're going to go back to multi-agent coordination patterns, which I've talked about before. I've talked about Cursor's idea that you have like a two-agent hierarchy, and that's much more effective than an agent swarm.”
The pattern of outcome specs + guardrails + tools + multi-agent coordination extends from current AI development infrastructure (factory.ai, Cursor) and is increasingly how large models should operate.
“We're just extending the current pattern we see at factory.ai, at Cursor, et cetera. And we're saying these models are more capable, you're going to give them more to work with.”
The first thing IT and security professionals should do when Mythos is released is to immediately battle-test it against their own systems and identify vulnerabilities—this is job number one, day zero.
“Which immediately suggests that the first thing you should do if you are in IT or security, as soon as Mythos comes out, is to say, 'Let's battle test it against our own systems, and let's see what vulnerabilities it finds.' That is That is job number one, day zero, you've got to do that.”
You can use current AI models right now (without waiting for Mythos) to find $200/month in household subscription savings by analyzing recurring charges and suggesting cancellations.
“You can do this now. You don't have to wait for Mythos. Just tell your AI, "Look at my recurring subscriptions for my household and find me $200 in savings." Most households in America can find $200 a month in savings somewhere from some of their subscriptions, just because the subscription economy latches onto our credit cards like barnacles, and we just need to find some space. And LLMs right now, without waiting for Claude Mythos, are already very good at that.”
End-of-pipeline evaluation scripts must comprehensively test functional requirements, non-functional requirements, exception handling, edge cases, and code quality—areas where humans currently serve as bottlenecks but must be automated.
“When you write the eval at the end, make sure that it tests absolutely everything. Your functional requirements, your non-functional requirements. And then make sure when it comes back that you are sure if you read that eval script and all the tests pass that you're confident it's good. Because increasingly, we need to be at a point where we're not the ones who are the bottleneck for telling whether a particular piece of software reliably calls dependencies. And whether the code is clean and has appropriate exception handling. Whether edge cases are handled. Like that all needs to go into automated evals”
For non-technical work, verification quality depends on maintaining a high personal standard—not accepting 99% correct work but insisting on fixing identified issues to prevent passing flawed outputs downstream.
“If you're doing non-technical work, the answer is going to be increasingly clear. You're going to look at it say, 'Wow, that looks really good.' and move on. And a lot of people are going to do that. And the art of checking non-technical work is increasingly going to involve you having a very high standard. One of the challenges of working with good models is learning to make sure our bar is really high. Don't be afraid just cuz this model is really good to look at something it produces like a PowerPoint deck or an Excel and say this isn't quite right. You got 99% of this right, but this is the 1% I want fixed.”
Outcome specifications may not be achievable today with current models, but organizations should start architecting systems to use outcome specs now to be ready for Mythos release.
“You may not be able to do that today with today's models, but be ready to do that. Get ready now because it's going to take time to re-architect your systems. Start to think about how you need to prioritize your work to get stuff.”
Tool definitions should clearly specify what each tool does so models can autonomously decide when to use them; effective tool definition is a design priority that improves with attention.
“Number three, think about the tools your model can use. You want to make sure you have an excellent set of tools that have the right capabilities where a model can look at the tool set and say, 'I know what this does. I can use it, and it's going to be effective at what it does.' So, right, maybe it's search your knowledge base, maybe it's look up an account, maybe it's process a refund. The model is going to decide increasingly what to call on in what order. And it's going to be up to you to define what the tools do, and to be effective about that, and present those tools effectively. So, put some work into your tool definitions.”
One Mythos model can assess the work of another Mythos model instance, providing evaluation without requiring human review, addressing the human bottleneck in code and artifact verification.
“you can also have a different instantiated Metheus model assess the work of that other model so that you don't have a second pair of eyes, right? You don't need to trust the model that did the work to do the eval as well.”
Example of good outcome specification for customer service: 'Resolve this customer's issue using our knowledge base, policies, and account history. Customer should leave satisfied and resolution comply with our return policy.' This contrasts with process-heavy specifications that route through intent classification, handler assignment, and specific retrieval methods.
“Let me give you an example back in the customer service world. Let's say what you want to specify is resolve this customer's issue using our knowledge base, our policies, and our account history, and you make sure the model has access to that. The customer should leave satisfied, and the resolution should comply with our return policy. That is actually a decent outcome specification. Compare that to what most production systems I see look like. It's essentially a process, right? Well, first classify the intent into one of 14 categories, then route to the appropriate handler, then retrieve the top five knowledge-based using hybrid search with alpha equals 0.7, then generate a response using only the retrieved context. That is what most of the prompts that I look at today look like.”
The checklist provided in the video (prompt simplification, retrieval architecture, hard-coded domain knowledge, output verification) should be worked through at personal and organizational level to prepare systems for Mythos.
“I hope this gives you a set of actionable recommendations to go with today to actually get ready for this. Because this is one of those things if you look at the checklist I gave you here in this video, you're going to need to do some work to get ready for this. If you want to be ready day one, your systems need to be ready to simplify. So, take the time whether it's your own personal [31:05] checklist and you're saying how what what what can prompts do I use? Can I simplify those? Or whether it's a work checklist and you're looking at like Agent systems and how you prepare for them and guardrails, get ready.”
The speaker previously used a 10-line research prompt that was actually over-constraining the model, and getting better results by reducing it to a one-liner, demonstrating that detailed process specification can harm rather than help with advanced models.
“I had an example recently where there was a a prompt that I was using around how I do research, and I'd been using it for a couple of model generations, and one day I forgot to put the full 10-line prompt in and I just put a one-liner and said go research and I got a better result back because the 10-line prompt...was more detailed about methodology than it needed to be and was over constraining my model.”
In software development, models now achieve 99% correctness more consistently than 85% correctness, fundamentally changing evaluation strategy and requiring comprehensive end-of-pipeline verification rather than intermediate checks.
“we are in a world now where we are closer to 99% right more of the time than we are to 85%. And Mythos is just going to push us more in that direction.”
Customer support agents with 3,000-token system prompts containing procedural steps (classify intent, check for hallucinated URLs, etc.) may be able to delete 30-50% of that prompt when a model becomes two to three times smarter without losing performance.
“Let's say you have a customer support agent with a 3,000 token system prompt, and half of it is procedural, right? Like, first classify the intent, then check your response for hallucinating URLs that do X, then do Y. That sequence was written down because the model would skip knowledge checks enough of the time that it was needed. Think about whether you're really going to need that when a model gets two or three times smarter. You may be able to delete that like 30, 40, 50% of that 3,000 token prompt because so much of the procedural is just not needed when the model is more intelligent.”
Models can infer house style for client reports at very high fidelity from a single example report, because the scaling law ensures that increased intelligence makes inference from examples more reliable.
“let's say you have a house style for writing your client reports. The model can just infer that at a very high fidelity from a given example report. That's the whole point of the scaling law is that intelligence gets better at reliably answering your query.”
Cybersecurity firm stock prices dropped 5-9% immediately upon Claude Mythos leaked details, indicating market recognition of the model's capability to make existing security approaches obsolete.
“we had another one of those stock market spook events where like security stocks for cybersecurity firms dropped between 5% and 9% as soon as Claude Mythos was even leaked. The entire world is starting to take AI models seriously because they have seen enough evidence to show what these models are capable of doing.”
The entire tech/investor community is taking AI models seriously now because they have seen sufficient evidence of capability, so claims about Mythos being a big jump in coding, artifact generation, reasoning, and cybersecurity should be taken seriously.
“The entire world is starting to take AI models seriously because they have seen enough evidence to show what these models are capable of doing.”
Claude Mythos will be expensive to run and likely initially available only to premium plan subscribers due to serving costs.
“these are not cheap models. Anthropic basically confirmed that. These are going to be very expensive models to run...I am willing to bet you that Mythos, when it launches, is only going to initially be available for Max plan users for Claude. Because it's so expensive to serve.”
Mythos will be much better at intelligently filling its context window than previous models because the scaling law shows that increased intelligence correlates with better context window usage, and this will continue the trajectory observed over the past 2 years.
“Do I have specific advice for you because Mythos is already out in the world and I've tested against 20 different things? I do not. We will get to that when it comes out. But I do have very, very high confidence that Mythos is going to be much better at this, much better at intelligently filling its context window than previous models because that is exactly what you get with the scaling law. When you get more intelligence, we we have seen over and over again, you get better at using your context window.”
Claude Mythos's release within weeks represents an inflection point requiring immediate preparation; organizations and individuals that prepare now will capture the train before it leaves the station.
“Claude Mythos is coming. The inflection point is here. This is another one of those moments when you need to be able to catch the train before it leaves the station. I hope this gives you a set of actionable recommendations to go with today to actually get ready for this. Because this is one of those things if you look at the checklist I gave you here in this video, you're going to need to do some work to get ready for this. If you want to be ready day one, your systems need to be ready to simplify.”
Current limitations in agentic software development include insufficient primitives for integration with external systems (like Google Calendar), but these limitations will be addressed in future model releases.
“Do we need better agentic primitives? Do we need better ways to hook into things like a Google calendar so it's easier? Absolutely. That is one of the big pain points right now, but that is the direction we're headed.”