2423.eth

CEO @Web3Rehashed // Curating Web3 narratives with an eye for art and asymmetry. @future // Commonware alum

[2423.eth @diaxbt]

Some serious stuff is happening on Ethereum. Spent a few days on my personal analysis and here what I've come up with. DYOR!➿

[RehashedDAO @Web3Rehashed]

ETHEREUM HAS SOLVED THE BLOCKCHAIN TRILEMMA.⁉️

Or, more accurately, @ethereum may have found a way around the assumption that made the trilemma feel unavoidable in the first place.

For years, blockchains were forced to choose between decentralization, security, and scalability. Push throughput too hard and node requirements rise. Keep nodes cheap and throughput suffers. Sacrifice decentralization and scaling suddenly becomes easy.

Ethereum's answer was not to pick one. It was to change the architecture entirely 👓

Article from @diaxbt:

I've always found the blockchain trilemma slightly misleading. Not because the tradeoff isn't real, but because it assumes every node has to process, download, and store everything the network does.

That is exactly the model Ethereum has been moving away from.

Instead of scaling one monolithic chain indefinitely, Ethereum separates execution, settlement, consensus, and data availability. Rollups execute activity, Ethereum settles it, consensus protects the canonical history, and the base layer provides the data needed for independent verification.

The first major step was rollups. Rather than putting every transaction directly onto Ethereum, L2s execute activity elsewhere and publish compressed results and data back to L1. That lets execution scale independently without requiring Ethereum itself to process every user transaction.

But rollups still had a bottleneck: data.

Originally, most rollup data was posted through calldata, which worked but was expensive. EIP-4844 changed this by introducing blobs, temporary data space designed specifically for rollups. Pectra later increased blob capacity further.

The problem was that every full node still had to download every blob. So simply increasing blob count forever would eventually recreate the same scalability problem.

This is where PeerDAS matters. 👽

PeerDAS, introduced with Fusaka, changes how Ethereum handles data availability. Instead of every node downloading all blob data, nodes only custody and sample portions of it. Erasure coding provides redundancy, allowing the full data to be reconstructed without requiring every participant to store everything.

That is the key shift.

Ethereum can increase total data availability capacity without increasing the burden on every individual node at the same rate.

This is fundamentally different from the traditional approach of simply making blocks larger and requiring stronger machines.

Longer term, full Danksharding extends the same model. Ethereum's roadmap targets a world where rollups can process far more activity while Ethereum remains the secure settlement and data availability layer underneath.

Verification is also becoming more efficient. Statelessness and zkEVM-based L1 verification aim to reduce how much local state and computation nodes need in order to independently verify Ethereum.

These pieces are not all finished, and that distinction matters. Full Danksharding is still ahead. Stateless Ethereum is unfinished. L1 zkEVM verification remains under development.

The L2 layer also has its own decentralization problems.

According to @l2beat, major rollups such as @base and @arbitrum still rely on upgrade mechanisms, sequencers, security councils, or other trust assumptions that keep them below fully mature Stage 2 decentralization.

So I would not interpret "Ethereum solved the trilemma" as meaning the tradeoffs disappeared.

What Ethereum appears to have solved is the architectural assumption behind them.

Scaling no longer has to mean making every node proportionally more powerful. Execution can move to rollups, data availability can be distributed across the network, and cryptographic proofs can reduce the cost of verification.

Ethereum stopped trying to make one blockchain do everything.

That may end up being the more important breakthrough.

[2423.eth @diaxbt]

@Web3Rehashed @AndrewTheiss Lfg!!!!

[2423.eth @diaxbt]

@aalexiamaga53

[2423.eth @diaxbt]

Some serious stuff is happening on Ethereum. Spent a few days on my personal analysis and here what I've come up with. DYOR!➿

[RehashedDAO @Web3Rehashed]

ETHEREUM HAS SOLVED THE BLOCKCHAIN TRILEMMA.⁉️

Or, more accurately, @ethereum may have found a way around the assumption that made the trilemma feel unavoidable in the first place.

For years, blockchains were forced to choose between decentralization, security, and scalability. Push throughput too hard and node requirements rise. Keep nodes cheap and throughput suffers. Sacrifice decentralization and scaling suddenly becomes easy.

Ethereum's answer was not to pick one. It was to change the architecture entirely 👓

Article from @diaxbt:

I've always found the blockchain trilemma slightly misleading. Not because the tradeoff isn't real, but because it assumes every node has to process, download, and store everything the network does.

That is exactly the model Ethereum has been moving away from.

Instead of scaling one monolithic chain indefinitely, Ethereum separates execution, settlement, consensus, and data availability. Rollups execute activity, Ethereum settles it, consensus protects the canonical history, and the base layer provides the data needed for independent verification.

The first major step was rollups. Rather than putting every transaction directly onto Ethereum, L2s execute activity elsewhere and publish compressed results and data back to L1. That lets execution scale independently without requiring Ethereum itself to process every user transaction.

But rollups still had a bottleneck: data.

Originally, most rollup data was posted through calldata, which worked but was expensive. EIP-4844 changed this by introducing blobs, temporary data space designed specifically for rollups. Pectra later increased blob capacity further.

The problem was that every full node still had to download every blob. So simply increasing blob count forever would eventually recreate the same scalability problem.

This is where PeerDAS matters. 👽

PeerDAS, introduced with Fusaka, changes how Ethereum handles data availability. Instead of every node downloading all blob data, nodes only custody and sample portions of it. Erasure coding provides redundancy, allowing the full data to be reconstructed without requiring every participant to store everything.

That is the key shift.

Ethereum can increase total data availability capacity without increasing the burden on every individual node at the same rate.

This is fundamentally different from the traditional approach of simply making blocks larger and requiring stronger machines.

Longer term, full Danksharding extends the same model. Ethereum's roadmap targets a world where rollups can process far more activity while Ethereum remains the secure settlement and data availability layer underneath.

Verification is also becoming more efficient. Statelessness and zkEVM-based L1 verification aim to reduce how much local state and computation nodes need in order to independently verify Ethereum.

These pieces are not all finished, and that distinction matters. Full Danksharding is still ahead. Stateless Ethereum is unfinished. L1 zkEVM verification remains under development.

The L2 layer also has its own decentralization problems.

According to @l2beat, major rollups such as @base and @arbitrum still rely on upgrade mechanisms, sequencers, security councils, or other trust assumptions that keep them below fully mature Stage 2 decentralization.

So I would not interpret "Ethereum solved the trilemma" as meaning the tradeoffs disappeared.

What Ethereum appears to have solved is the architectural assumption behind them.

Scaling no longer has to mean making every node proportionally more powerful. Execution can move to rollups, data availability can be distributed across the network, and cryptographic proofs can reduce the cost of verification.

Ethereum stopped trying to make one blockchain do everything.

That may end up being the more important breakthrough.

[2423.eth @diaxbt]

@Ellymams Thx sweetheart!

[2423.eth @diaxbt]

@shittu_a33260 True

[2423.eth @diaxbt]

@gevin7777

[2423.eth @diaxbt]

Enjoyed this one! Thx to Jim for sharing his thoughts on some seriously under-discussed topics.👏

[RehashedDAO @Web3Rehashed]

Are we really heading toward a million-chain future?

@jimchang shares his view on rollups, RaaS, data availability, and how modular infrastructure is making it easier to build chains around very specific goals.

[RehashedDAO @Web3Rehashed]

What actually makes a blockchain a blockchain?

In recent episode of Rehashed, Jimmy Chang (@jimchang) breaks down the difference between monolithic chains, modular chains, and DAGs, then walks through the 4 core pillars behind blockchain architecture: execution, transaction ordering, settlement, and data availability.

A proper deep dive into how modern chains are actually designed.

[Pixel Symphony @Pixel0Symphony]

Franco Grignani, Unknown. 1965.

[RehashedDAO @Web3Rehashed]

Luca’s mission is ultimately straightforward: show the world the greatness of @pudgypenguins.

Not as a speculative asset, and not as an artifact of one particular market cycle, but as a globally recognisable universe with characters, products, stories, and a community people genuinely want to belong to.

That is the deeper test for crypto-native brands: can they turn attention into affinity, and affinity into something durable?

With Pudgy World and Overpass on the horizon, Pudgy Penguins is continuing to build toward that vision.

[RehashedDAO @Web3Rehashed]

For Luca, the ambition has never been to make Pudgy Penguins successful only within crypto.🐧

The larger objective is to build a mass-market brand, one that can reach people who have never traded an NFT, used a wallet, or followed the culture that produced the project. That means meeting people where they already are: through products, retail, characters, entertainment, and experiences that do not require technical literacy to understand.

[RehashedDAO @Web3Rehashed]

On this episode of @Web3Rehashed, we speak with @LucaNetz, CEO of @pudgypenguins, about one of crypto’s most difficult questions: how does an internet-native project become a brand that can survive beyond the cycle that created it?

Pudgy Penguins launched on Ethereum in July 2021. In April 2022, Luca acquired the project for 750 ETH, then set out to transform it from a well-known NFT collection into a broader consumer-facing company.

Today, Pudgy products sit in Walmart stores across the United States, an outcome that would have sounded improbable for a PFP project only a few years ago.

This conversation explores the decisions, risks, and long-term thinking behind that transition.

[2423.eth @diaxbt]

@Web3Rehashed @LucaNetz @pudgypenguins

[2423.eth @diaxbt]

@Web3Rehashed @HesterPeirce Great episode, fam!

[RehashedDAO @Web3Rehashed]

Revisiting our long-awaited conversation with SEC Commissioner @HesterPeirce.

For years, Commissioner Peirce has remained one of the most intellectually rigorous and independent voices within the SEC on questions of digital assets.

She has repeatedly challenged the tendency to treat technological experimentation as a regulatory threat by default, arguing instead for clearer standards, proportionate oversight, and a framework that does not punish innovation through ambiguity.

Her perspective remains essential for anyone trying to understand the institutional tensions shaping crypto policy in the United States.

[RehashedDAO @Web3Rehashed]

The SEC is often perceived as distant, opaque, and fundamentally inaccessible to early-stage builders. That perception is understandable, but it is not inevitable.

In our conversation with SEC Commissioner @HesterPeirce on @web3Rehashed, she offered a more practical view of regulatory engagement: approach the institution early, communicate with precision, and treat compliance not as an afterthought, but as part of building durable infrastructure.

For founders operating in an uncertain regulatory environment, that distinction can be decisive.

[RehashedDAO @Web3Rehashed]

The core news is not that @AnthropicAI temporarily took two models offline.

It is that the United States has now treated access to frontier AI as something that can be controlled through export law, not only by geography, but by who the user is.

Anthropic was reportedly ordered to suspend access to Fable 5 and Mythos 5 for any foreign national, including people physically located inside the United States and even foreign-national Anthropic employees.

That is a very different regime from blocking chip shipments to China or limiting GPU clusters abroad.

It is closer to saying: the intelligence itself has become a controlled strategic asset. 👇

~~ Analysis by @onchainhost ~~

For years, the AI race has been framed around inputs.

Who can buy the best GPUs. Who can secure HBM supply. Who can build data centers fast enough. Who can access advanced lithography, power capacity, and cloud credits.

The US export-control strategy reflected that logic: restrict Nvidia-class hardware, constrain semiconductor tooling, limit advanced compute in adversarial jurisdictions, and slow down the ability to train frontier models.

That framework made intuitive sense because chips are physical. They cross borders. They are manufactured, shipped, tracked, and sold through identifiable supply chains.

But models are different.

A frontier model can sit in a US data center while serving someone on the other side of the world through an API. Nothing physical crosses a border. No GPU is exported. No model weights need to leave the company. A few API calls can still deliver capabilities that used to require access to an entire research organization.

That is why this @AnthropicAI episode matters.

The Commerce Department directive reportedly required a license for foreign persons to access Fable 5 and Mythos 5, regardless of whether those users were in the US or abroad. Anthropic said it could not reliably separate its users by nationality, so the practical outcome was to disable both models for everyone.

In other words, the state did not control the hardware.

It controlled the ability to query the intelligence running on the hardware.

That distinction may sound technical, but it changes the structure of the AI market.

A frontier API is no longer simply a cloud product.

It can become a licensed strategic service.

Anthropic’s Fable 5 had been released as a generally accessible model with cybersecurity safeguards. Mythos 5 was more restricted, intended for a smaller trusted-access group where some cyber safeguards were lifted for defensive use cases.

Anthropic itself described Mythos-class systems as a higher capability tier than its Opus models, particularly for software engineering, autonomous work, cyber defense, and scientific research.

The government’s concern was reportedly that Fable’s safeguards could be bypassed in a way that enabled users to identify software vulnerabilities.

Anthropic disputes the characterization. The company says the technique was narrow, non-universal, involved a limited number of previously known minor vulnerabilities, and did not demonstrate a capability unique to its models.

It also argues that similar bug-finding behavior is available through other public models.

The truth is that both sides may be describing a real problem from different angles.

Anthropic is right that jailbreak resistance is not binary. A model can have strong protections and still be vulnerable in narrow contexts. That is the nature of frontier model security today: safeguards reduce the cost of defense, but they do not produce perfect containment.

The government is also right about one thing: capability diffusion does not need to be perfect to matter.

A model does not need to autonomously compromise a military network to create strategic risk. It may be enough for it to make skilled researchers, cyber operators, or intelligence teams materially faster at vulnerability discovery, exploit research, systems analysis, or code review.

The issue is therefore not whether a model is “dangerous” in the abstract.

The issue is whether certain increments of capability are significant enough that access itself becomes a national-security question.

That is a much harder line to draw than the chip line.

A chip can be classified by performance thresholds. Compute capacity can be estimated. Interconnect bandwidth can be measured. A model’s strategic value is more contextual.

The same model that helps a defensive security team patch an aging banking system can help an offensive researcher find weaknesses in that system.

The same agent that compresses months of software engineering into days can compress reconnaissance, reverse engineering, and exploit development.

And the same model that can be used by a US cybersecurity firm through a legitimate API can potentially be used by a foreign actor through the exact same interface.

This is why the “foreign national” language is the most consequential part of the story.

The policy is not simply saying: do not serve sanctioned jurisdictions.

It is applying the logic of deemed exports, where releasing controlled technology to a foreign person inside the United States can be treated as an export to that person’s country of nationality.

That principle already exists in traditional export controls. What is new is applying it to real-time access to a commercial frontier model.

This makes the situation less like a normal product restriction and more like an emergency intervention.

And that uncertainty is itself a market signal.

For enterprises building core workflows around frontier APIs, the risk is no longer limited to pricing changes, rate limits, outages, or model deprecation.

There is now geopolitical dependency risk.

A company in London, Seoul, Dubai, Singapore, or Istanbul can build its product architecture around a US model, integrate it deeply into engineering workflows, and then discover that access is conditional on a political or regulatory decision made in Washington.

That is not a theoretical concern anymore.

Anthropic’s own decision to disable access globally shows the operational reality. A compliance requirement aimed at foreign nationals became, in practice, a kill switch for everyone because identity verification, citizenship classification, licensing, corporate ownership analysis, and access enforcement are extremely difficult to implement across global cloud infrastructure.

This is where the AI sovereignty conversation becomes much more concrete.

For a long time, “sovereign AI” sounded like a policy slogan: countries wanting local language models, domestic clusters, national compute programs, or regional data residency.

Now it has a more practical meaning.

Sovereignty is not only about owning GPUs.

It is about whether a government, company, university, security team, or startup can maintain access to the intelligence layer when geopolitical conditions change.

That will make open-weight models more strategically attractive, even when they are less capable.

Not necessarily because open models are better.

But because a model that can run on infrastructure you control cannot be switched off by a foreign provider under an emergency licensing directive.

That creates a major tradeoff.

Closed frontier systems may remain ahead in capability, reliability, tool use, long-horizon reasoning, and safety infrastructure. But they also concentrate political and regulatory power inside the provider’s jurisdiction.

Open-weight systems sacrifice some of that frontier performance, but they reduce dependence on a single company, a single cloud platform, or a single national export-control regime.

For builders, this probably accelerates a multi-model future.

The question will not only be: “Which model performs best?”

It will increasingly be: “Which model can we still access under stress?”

That could mean enterprises keeping secondary model providers, designing agent stacks that can swap inference backends, maintaining open-weight fallback systems, or avoiding architecture that assumes one frontier API will always be globally available.

This does not mean the US will suddenly export-control every powerful model.

The current action is specific to Anthropic’s Fable 5 and Mythos 5, and the facts remain contested. The government has not publicly released the full legal reasoning, while Anthropic maintains that the cited jailbreak was narrow and that broader restrictions would risk halting frontier deployment across the industry.

Still, precedents matter more than permanence.

Once a government demonstrates that it can regulate model access through export-control authorities, every other frontier lab has to plan around that possibility.

@OpenAI, @GoogleDeepMind, @xai, @Meta, and the major cloud providers all now have a reason to ask the same uncomfortable question:

At what level of capability does a model stop being software and start being controlled strategic infrastructure?

The answer will shape more than AI policy.

It will shape where startups build, how enterprises procure models, why countries invest in domestic compute, and whether the next generation of AI becomes globally accessible infrastructure or a fragmented network of national capability zones.

The most important thing to watch is not whether Fable 5 comes back online next week.

It is whether this becomes a one-off dispute around @AnthropicAI, or the first real template for governing frontier intelligence as an export-controlled resource.

[2423.eth @diaxbt]

Billy is a true master in this space. What he’s building at Intuition feels genuinely exciting, such a powerful episode.

But honestly, it also stirred something in me. I’ve been talking about all of this with my husband, wondering what happens when AI starts handling all the basic parts of life. What if the world slowly becomes… plastic? Smooth, efficient, but hollow. Empty of real emotion.

[RehashedDAO @Web3Rehashed]

At one point in our episode with @0xbilly, we went down a bit of a dystopian rabbit hole.

The conversation started with creativity, AI, and the speed at which ideas can now become real, but quickly moved into a bigger question: what happens when the world around us becomes easier to generate, manipulate, and personalize?

Billy reflected on a future where AI does not just help us create faster, but begins to reshape how we experience reality itself. The line between imagination, media, identity, and environment starts getting thinner, and that creates both incredible possibilities and very real risks.

[RehashedDAO @Web3Rehashed]

Two days ago, Anthropic launched Claude Fable 5 and Claude Mythos 5. On the surface, this looks like another frontier model release. But I think the more important story is not just capability.

It is access.

Fable 5 is the public Mythos-class model. Mythos 5 is the same underlying model, but with some safeguards lifted for trusted cyber defenders, infrastructure providers, and eventually select biology researchers.

In its early stages, Anthropic is demonstrating what the next phase of frontier AI deployment may look like👇

~~ Analysis by @punkbennet ~~

I, like many others, have become slightly numb to model launches.

Every few months, a new model arrives with better coding, better reasoning, better long-context performance, better benchmark charts, better agentic workflows, better everything. At some point, the launch cycle starts to blur into one long benchmark war.

But the Fable / Mythos release feels different.

Not because Anthropic is claiming another step forward in intelligence, though it is. Not because the model seems strong at long-horizon coding, scientific reasoning, vision, finance, and complex knowledge work, though that matters too.

It feels different because Anthropic is openly splitting capability into two layers: a public version and a trusted-access version.

That is the real story.

Fable 5 is described as a Mythos-class model made safe for general use. According to Anthropic, it exceeds any model they have previously made generally available and is especially strong on long, complex tasks. The model is available to general users, but it ships with classifiers that detect certain categories of high-risk use.

When those classifiers trigger, the request does not get handled by Fable 5. It falls back to Claude Opus 4.8.

The covered areas are cybersecurity, biology and chemistry, and distillation. In plain language: domains where the model’s raw capability could create meaningful risk if used badly, or where Anthropic believes unrestricted access could accelerate misuse or capability proliferation.

Anthropic says these safeguards trigger in less than 5% of sessions on average, meaning most users should experience Fable as the full Mythos-class model most of the time. But that 5% is where the entire debate lives.

Because if you are building a normal app, analyzing documents, writing code, doing finance work, or working on general research, Fable 5 may simply feel like a stronger frontier model.

If you are doing security research, advanced biology, chemistry, or frontier model development, the product experience becomes more complicated.

That is where Mythos 5 comes in.

Mythos 5 is the same underlying model as Fable 5, but with safeguards lifted in some areas. It is not generally available. It is being deployed through Project Glasswing, Anthropic’s initiative with cyber defenders and critical software infrastructure providers. Anthropic says it plans to expand access through a broader trusted program.

This is a meaningful shift from “everyone gets the same model” to “capability access depends on trust, use case, and risk category.”

I do not think this is just product packaging.

It is probably a preview of how frontier AI gets distributed from here.

In crypto, we are used to open access as a cultural default. The whole industry is built around permissionless infrastructure, public networks, open liquidity, composability, and adversarial testing. The assumption is that if something is powerful, the network should expose it, and the market should figure out what survives.

Frontier AI is moving in a different direction.

The most capable systems are becoming too useful to keep entirely closed, but too risky to release without restrictions. That creates a middle layer: broad public access for most tasks, gated access for sensitive domains, and institutional partnerships for the highest-risk capabilities.

There is a strong argument for this.

If a model is genuinely good at finding and exploiting software vulnerabilities, then unrestricted release has obvious downside. Anthropic previously said Mythos Preview had found thousands of high-severity vulnerabilities, including some in major operating systems and browsers. Even if we treat that as an Anthropic claim rather than independent proof, the direction is clear: models are becoming serious cyber tools.

That means the same capability can be defensive or offensive depending on who holds it.

A security team using Mythos to audit critical infrastructure is very different from an unknown actor using it to automate exploit discovery. A biology researcher using the model to generate therapeutic hypotheses is very different from someone trying to gain dangerous biological uplift.

The difficult part is that the boundary is not clean.

Dual-use work is messy. Real security research can look like offensive security. Real biology can overlap with sensitive methods. Real AI research can look like distillation or capability extraction. If the classifier is too narrow, malicious users get through. If it is too broad, legitimate researchers get blocked or silently downgraded.

This is why the transparency issue matters.

After launch, Anthropic already faced backlash around invisible safeguards for frontier LLM development. The criticism was not only that the model had restrictions. Most serious users understand that frontier systems will have restrictions. The criticism was that some interventions were not visible enough to the user, which makes evaluation harder and damages trust.

If a model refuses, that is annoying but clear.

If a model falls back to a weaker model and tells you, that is also clear.

But if a model quietly changes behavior, limits effectiveness, or routes around your task without making the intervention obvious, then developers cannot properly evaluate it. Researchers cannot know whether they are testing model capability, product policy, or invisible steering.

That is a major problem.

To Anthropic’s credit, they appear to have recognized this quickly and said they are changing Fable 5’s safeguards for frontier LLM development to make them visible. That is the right direction.

Still, the tension does not disappear.

The bigger question is whether frontier AI companies can build trust while also reserving the most powerful capabilities for trusted actors. This is not just about Anthropic. It is about the governance model of the entire AI stack.

The public wants access.

Developers want predictable behavior.

Researchers want measurable capability.

Governments want security.

Labs want to avoid catastrophic misuse.

Competitors want fair evaluation.

Enterprises want privacy, reliability, and compliance.

All of those demands collide inside a release like Fable / Mythos.

Another under-discussed piece is data retention. Anthropic says Mythos-class traffic requires 30-day retention for safety monitoring, while also saying the data will not be used to train new Claude models and will be deleted after 30 days in almost all cases.

That may be reasonable from a safety perspective, especially if the goal is detecting jailbreaks or coordinated misuse across many requests.

But for enterprises, regulated industries, and sensitive research teams, it becomes a real deployment consideration. The more capable the model, the more likely users want to use it on sensitive work. The more sensitive the work, the more important retention policy becomes.

So the model is not just competing on intelligence anymore.

It is competing on governance.

This is probably where the AI market is going. The best model will not simply be the one with the highest benchmark score. It will be the one that offers the best combination of capability, transparency, access control, reliability, compliance, cost, and trust.

Fable 5 and Mythos 5 are interesting because they expose that full stack at once.

There is the capability story: a model above Opus-class, built for long-horizon tasks and advanced reasoning.

There is the safety story: classifiers, fallbacks, red-teaming, limited access, and trusted programs.

There is the product story: public users get Fable, vetted users get Mythos.

There is the trust story: users need to know when they are interacting with full capability and when safeguards are shaping the output.

There is the market story: frontier AI is becoming less like a normal SaaS product and more like critical infrastructure.

Personally, I think this release is one of the clearest signs that “open vs closed” is no longer the only useful framing.

The new framing is closer to: who gets which capability, under what conditions, with what monitoring, and with what disclosure?

That is less clean than the old debate, but probably more accurate.

Based on the available information, Fable 5 may become an important public frontier model. Mythos 5 may become an important restricted capability layer for security and science. But the bigger experiment is the access model itself.

If Anthropic gets the balance right, this could become a template for deploying very powerful AI safely while still letting most users benefit from the capability.

If they get it wrong, it becomes a trust problem: too much opacity for developers, too much restriction for researchers, and too much central control over frontier capability.

Either way, this is worth watching.

Not just because Mythos looks powerful.

Because it shows how AI labs may decide who is allowed to use power at all.