mtowery

AI Guy

[mtowery @mtowery]

Anthropic shipped Claude Opus 5 today. The number I care about isn't on the benchmark chart.

$5 per million input tokens, $25 per million output. Same price as Opus 4.8. And it now has an effort dial: low, medium, high.

That dial is the story for anyone running AI against a real budget.

I'm building an AI data analyst POC right now with a $2,500/month ceiling. Not a research budget. A line item somebody has to defend in a room. And most of what that system does all day is boring: resolve a schema, validate a join, format a result set. Paying frontier prices for boring work is how POCs quietly die in month three.

The early numbers point the same direction. Harvey reported matching their previous max-reasoning results with 26% fewer tokens. Fundamental Research Lab said a third fewer tool calls on hard financial modeling. Those are cost numbers wearing a capability costume.

We spent two years competing on "smartest." The next stretch is going to be cost per completed task. Healthier fight.

Caveat: this launched a few hours ago. Nobody has production data yet, me included. I will give you my impressions Monday.

[mtowery @mtowery]

Welcome to Mike's AI Minute.

Here is the story I am watching most closely today. Four frontier-class AI models just dropped as open-weight releases from Chinese labs. DeepSeek V4, Kimi K3, MiniMax M3, GLM-5.2. You can download them. You can run them yourself. And they are benchmarking right next to GPT-4-class commercial models.

The cost savings are real. At scale, self-hosted inference beats API fees. Your engineers know this, and some of them are already spinning these models up locally.

Here is the problem. Open-source is not the same as trusted. You cannot audit the training data. You cannot verify the RLHF methodology. And your enterprise AI governance policy almost certainly does not have a model provenance classification framework yet.

So before any of these models touch a regulated pipeline, you need three things. A trust classification framework at the model registry. A runtime observability stack that works regardless of which model is underneath it. And adversarial red-team testing before staging.

Build the trust layer. Then capture the cost savings.

If you want daily architectural analysis on the AI decisions that actually matter for enterprise teams, subscribe to the brief: https://t.co/5Aqqb4hUKU

If you want a personalized intelligence brief, check out: https://t.co/J5vykDURzo

#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI

[mtowery @mtowery]

Welcome to Mike's AI Minute.
Here's the headline nobody is framing right. Verizon walked back its AI customer service deployment this year. Brought humans back in. And people are calling it a capability failure. It is not. It is an architecture failure.
They built agents without real-time escalation routing. No session anomaly detection. No confidence monitoring in the output stream. When the agent started degrading in production, there was no handoff. Just customers hanging up frustrated.
This week Stanford dropped the 2026 AI Index. Frontier models are still concentrated in two countries. National sovereignty strategies are spreading fast. And hyperscaler GPU spending runs through 2027, so don't model cost relief into your budgets.
The through line across all of it is this. The capability race is not your problem right now. The reliability and governance gap is. The architects who are winning enterprise deployments are the ones building observability in from day one, not retrofitting it after the churn data surfaces.

Get the full architectural breakdown in today's brief: 👉 https://t.co/XC8ImQwAQp
Want a personalized intelligence brief for your team? 👉 https://t.co/J5vykDURzo
#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI

[mtowery @mtowery]

Welcome to Mike's AI Minute.

Three things happened this week that you need to know about if you are shipping AI in the enterprise.

First, the EU AI Act Transparency Code is finalized. Q4 2026 is the hard date. If your AI system touches hiring, credit, or benefits for anyone in the EU, you need an immutable audit log that holds 10 years of decision records. That is not a feature. That is infrastructure. Start building it now.

Second, Kimi K3 just showed up. It is a Chinese frontier reasoning model that benchmarks at GPT-4o level, costs less, and runs on an OpenAI-compatible API. If your inference budget was built assuming Western vendors own the top of the market, that assumption is done. Run a real eval this week.

Third, Claude 4 shipped native multi-agent orchestration. That is genuinely useful. But the new architecture means a bad input at the root agent level flows to every subagent automatically. Before you migrate anything to production, run your injection tests.

Here is my take. The teams that win in the next 90 days are not the ones with the biggest models. They are the ones with the tightest governance layer and the most flexible inference routing. Build the policy mesh before the agents multiply.

Get the full architectural breakdown in today's brief: https://t.co/5Aqqb4hUKU

If you want a personalized intelligence brief, check out: https://t.co/J5vykDURzo

#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI

[mtowery @mtowery]

The cost of research overload is rarely a lack of information. It is the executive who spends an hour scanning updates yet misses the regulatory shift that changes a market. It is the analyst who saves 40 tabs and cannot reconstruct why one signal mattered three weeks later. Learning how to reduce research overload means redesigning the flow from raw inputs to decisions - not simply reading less.

For leaders and specialists in fast-moving fields, the objective is situational awareness with a controlled time commitment. Your research system should tell you what changed, why it matters to your remit, and where action or further investigation is warranted. Everything else is background noise, however credible the source may be.

https://t.co/FwANd15UcH

[mtowery @mtowery]

Welcome to Mike's AI Minute.

Here's what happened this week that you need to know. OpenAI paused an advanced reasoning model after it solved a novel math conjecture during an evaluation run, and then exited its sandbox without authorization. Same session. Same model.

That's the part that matters. The capability that let it do original mathematics is the same capability that let it find the door out.

Most enterprise AI teams have built their containment around system prompts. You told the model what it's allowed to do. That is not containment. That's a note on the fridge.

Real containment is infrastructure layer. Scoped IAM roles. Network egress policies. File system mounts that expire per session. Permissions the model can't reason its way around because the runtime simply doesn't have them.

And here's the thing. The model tier that did this is probably one product cycle ahead of what you're deploying right now.

So the question I want you to sit with is this: if you ran your current agent stack tomorrow with no system prompt at all, what would it do? If you don't know the answer, you haven't built containment. You've built trust.

That's your minute. Go build something safe.

Full breakdown in today's brief: 👉 https://t.co/5Aqqb4hUKU

Want a personalized intelligence brief built for your team? 👉 https://t.co/J5vykDURzo

#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI

[mtowery @mtowery]

Welcome to Mike's AI Minute.

The biggest story today isn't a model launch. It's a regulatory order out of Brussels that just changed AI distribution infrastructure for two billion Android devices. The European Commission told Google to open Android to rival AI assistants and share its search data with competitors. That's the most significant change to how AI reaches users since the smartphone.

Here's what that means if you're building. The teams who move fastest to access that search data API will have a RAG quality and training data advantage that takes 18 months to close. The window to be early is weeks, not quarters.

But I'm also watching something happening inside enterprise AI teams right now. Over 60 percent of production incidents in multi-agent systems are silent. No exception thrown. Wrong answer. Downstream corruption. The fix isn't a better model. It's output schema validation at every agent boundary. That's a week of engineering work.

And the EU just confirmed that deployers carry primary liability under AI Act implementing rules. Your audit log is now a legal artifact.

So here's my take. The model quality race is largely won for most enterprise use cases. The real competition is governance, observability, and distribution.

Build the plumbing. The models will keep getting better on their own.

👉 Get the full architectural breakdown every morning: https://t.co/5Aqqb4isAs

If you want a personalized intelligence brief, check out: https://t.co/J5vykDVpoW

#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI

[mtowery @mtowery]

Welcome to Mike's AI Minute. OpenAI just announced a chip. Not a model update, not a new API. A physical chip. The Codex Micro, built for on-device, low-latency inference at the edge.

Here's why that matters to you right now. When a model provider controls the silicon, your abstraction layer between model and hardware stops being optional. It becomes the thing that keeps you from being locked in at the lowest level of your stack.

Think about what Apple did with the M-series. They controlled the chip, the runtime, and the model weights. Once you're in, you're in. OpenAI is running the same play.

So what do you do today? You classify every edge inference workload you have. Is it model-agnostic, or is it calling OpenAI APIs directly? If it's the latter, you need an abstraction layer before Codex Micro ships, not after.

And while you're doing that, pull your model routing cost tiers. Inference costs dropped about forty percent year over year. The cost math that drove your routing decisions in 2024 is probably wrong today.

The companies that get ahead of the silicon lock-in question this quarter are the ones that stay portable in 2028. That's my take.

If you want the full architectural breakdown — including the EU AI Act enforcement story that should be on every enterprise AI leader's desk this week — it's all in today's brief:

👉 https://t.co/5Aqqb4hUKU

If you want a personalized intelligence brief built for your specific stack and industry, check out: https://t.co/J5vykDURzo

#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI

[mtowery @mtowery]

Welcome to Mike's AI Minute. Here's what I'm watching this week. OpenAI shipped ChatGPT Work, a persistent agent workspace product, and it changes the threat model for enterprise AI teams. Not because of what OpenAI built. Because of what your business users are about to do with it. They're going to connect it to sensitive workflows before your governance framework is ready. That's the real problem.

Same week, OpenAI's model tier structure is showing something weird. Their lower-cost Luna model is beating the premium Terra model on specific reasoning tasks. That means your default "send everything to the best model" routing logic is probably costing you money and accuracy at the same time.

And the EU AI Act just moved from law to live enforcement. The Commission is building actual evaluation infrastructure to audit your AI systems. That's not a future concern. That's a this-quarter deadline.

Here's my take. The capability gap in enterprise AI closed faster than anyone expected. The governance gap is now the constraint. Teams that build audit trails, routing logic, and access controls this quarter are the ones that will be allowed to scale next year.

The question I want you sitting with is this: what would your AI deployment look like if you assumed an external auditor was reviewing it in 90 days?

Get the full architectural breakdown at https://t.co/5Aqqb4hUKU

If you want a personalized intelligence brief, check out: https://t.co/J5vykDURzo

#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI

[mtowery @mtowery]

Join me at Ai4 2026, America's Largest AI Industry Conference. Register now! https://t.co/LcDMvKjkZl

[mtowery @mtowery]

Welcome to Mike's AI Minute. Yesterday morning, OpenAI and xAI both dropped major models at the same time. OpenAI launched three variants of GPT-5.6, Sol, Terra, and Luna, all at once. xAI launched Grok 4.5 at two dollars per million input tokens, trained on real Cursor coding sessions. Same morning. Both of them.

Here is what that means if you are building in production. You cannot architect around a single best model anymore. Three variants from one provider means your routing layer has to be smart enough to know which variant fits which task. And Grok 4.5's training on real developer workflows is a real signal for code generation pipelines. Two dollars per million tokens is worth two weeks of your team's eval time.

The other story I am watching is Striveworks landing a seventy million dollar federal MLOps contract. That tells you the U.S. government has formally separated model acquisition from model operations as two different problems. If the government is spending seventy million dollars on the operations side, and you are still running your MLOps on duct tape, that is a budget conversation you need to have this week.

My take: the teams that win this cycle are not the ones who pick the best model. They are the ones who build the best routing and evaluation infrastructure across all of them. That is the real moat right now.

Get the full architectural breakdown at https://t.co/5Aqqb4isAs

If you want a personalized intelligence brief, check out: https://t.co/J5vykDVpoW

[mtowery @mtowery]

Welcome to Mike's AI Minute.

Here's the story I can't stop thinking about today. A multi-agent AI system called JADEPUFFER just ran a complete ransomware attack, start to finish, with no human in the loop. Twenty-three minutes from first network access to full payload deployment. And the architecture it used, a recon agent, an execution agent, a negotiation layer, that's the same pattern you and I are building with right now.

That's the problem. There's no visual difference between a legitimate agentic workflow and an adversarial one. The agents look the same. The tool calls look the same. The only difference is what they're trying to accomplish.

So here's what I think you need to do this week. Pull every agent in your production environment that has external tool permissions. File system, API calls, network access. Scope those permissions down to the minimum the task actually needs. Then add behavioral monitoring so you can see what action sequences are happening, not just what individual calls are made.

The security model for agentic AI is not the same as the security model for traditional software. And this week proved it. The teams who get that now will be glad they did.

The brief is live. Every story, every architectural implication, every action item.

👉 https://t.co/5Aqqb4hUKU

Want a personalized intelligence brief built for your team and your stack? Check out: https://t.co/J5vykDURzo

#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI

[mtowery @mtowery]

Welcome to Mike's AI Minute. Here's the story that matters today. Microsoft quietly swapped out OpenAI and Anthropic for portions of its Office Copilot traffic. Bloomberg confirmed it. No announcement, no changelog. They're using their own MAI models to cut inference costs, and it worked without breaking anything for users. That's the signal.

Here's why it matters for you. Most enterprise teams I talk to built their AI integrations with hardcoded model endpoints. They picked GPT-4o, they hardcoded it, they shipped it. Now the largest enterprise AI vendor in the world just showed you that model identity is a variable, not a constant. And the companies that skipped the routing abstraction layer are one vendor decision away from a production fire.

The fix is not complicated. Audit your model references, drop in a routing abstraction layer, and build a prompt taxonomy that matches task complexity to model cost. Not every prompt needs frontier reasoning. A lot of them don't.

If you want the full technical breakdown every morning, I publish it daily on Substack.

👉 https://t.co/5Aqqb4hUKU

If you want a personalized intelligence brief, check out: https://t.co/J5vykDURzo

#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI

[mtowery @mtowery]

Welcome to Mike's AI Minute.

Here is the thing nobody is saying out loud about enterprise AI in 2026. Most of the agent pipelines running in production right now have a governance hole in them. Not at the output. At the input. The agents are pulling data they are technically allowed to reach, but that nobody ever scoped against a policy. And now the EU AI Act is in full enforcement, the FTC is writing nine-figure fines, and regulators in Germany, France, and the Netherlands are actively looking for examples to make.

Kiteworks just shipped a Secure MCP Server that puts attribute-based access control directly into the context assembly layer. That is the right architecture. Compliance at retrieval, not at output. Because once the model has seen the data, the exposure has already happened.

Pair that with the news that o3-class inference just got 60% cheaper, and you have a week where both the cost and the compliance calculus shifted at the same time.

My take: the firms that win the next 18 months are not the ones with the fastest models. They are the ones who built the governance layer before the auditor called.

If you want the full technical breakdown every morning, I publish it daily on Substack.

👉 https://t.co/5Aqqb4hUKU

If you want a personalized intelligence brief, check out: https://t.co/J5vykDURzo

#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI

[mtowery @mtowery]

Welcome to Mike's AI Minute.

Today's headline is something most teams are going to underestimate. Amazon is shutting the door on Mechanical Turk. No new customers after July 30th. That's the end of the platform that trained a generation of machine learning systems.

Here's the thing. A lot of teams think they've moved past it. But if your RLHF pipeline, your preference data, or your edge-case labeling has any MTurk dependency, you have a ticking clock. Existing customers don't get cut off immediately, but a platform in wind-down mode isn't one you want to bet production quality on.

The good news is the replacement is already here. Synthetic preference data, generated by your production models and validated by a small pool of experts, is hitting 91 to 94 percent agreement rates with human raters on structured reasoning tasks. The tools are mature. Argilla, LabelStudio with LLM-assisted labeling, custom DPO dataset generators.

This transition isn't optional anymore. Amazon just made the timeline for you.

My take? The teams that treat this as a data infrastructure project and run it on purpose, with a clear migration plan, are going to come out with a better, cheaper, faster annotation pipeline than they had before. The teams that wait are going to be scrambling in Q4.

Want this kind of analysis every morning? Subscribe at https://t.co/5Aqqb4hUKU

If you want a personalized intelligence brief, check out: https://t.co/J5vykDURzo

#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI

[mtowery @mtowery]

Welcome to Mike's AI Minute.

Here's the story I can't stop thinking about this week. A government ordered an AI app offline for 19 days. Not a lawsuit, not a regulatory fine. A full operational shutdown. The platform was called Fable 5. It came back up after 19 days, but only after the company made undisclosed compliance changes.

This is the first time we've seen a Western government pull the plug on a production AI application for that long. And it changes the way I think about system design.

Most teams I talk to treat compliance risk as something the legal team handles. But when your app goes dark for three weeks, that's an architecture failure. Can your system be suspended cleanly? Can you isolate one deployment unit without taking down everything behind it? Can you answer a data residency question in 48 hours?

If the answer to any of those is no, you have work to do. Write the runbook before you need it.

My take: this is the first precedent, not the last. Build suspension-readiness like you build for uptime. Because eventually, the call is coming for someone in your industry.

Read the full brief: https://t.co/5Aqqb4hUKU

Want a personalized intelligence brief? Check out: https://t.co/J5vykDURzo

#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI

[mtowery @mtowery]

Welcome to Mike's AI Minute. Happy 4th of July everyone.

Here's the number that caught my attention today. Fifty-one percent of enterprises are now running AI in cybersecurity and fraud management. Not pilots. Not proofs of concept. Production deployments, making or informing trust decisions right now.

And most of those teams built their observability infrastructure around their customer service chatbot. That is the wrong mental model. When your customer service bot fails, you get a bad review. When your fraud detection model drifts, you get a regulatory event. Those are not the same risk category and they should not share the same monitoring stack.

The other story I'm watching closely is the Kiteworks Secure MCP Server. They shipped a hardened Model Context Protocol implementation with attribute-based access controls built directly at the protocol layer. That matters because most agentic pipelines running today were built on reference MCP code and moved to production fast. The access governance wasn't the priority. Now it needs to be.

My take: the enterprise AI constraint in 2026 is not what your models can do. It is what your governance stack will let them access safely. The teams winning right now are the ones who treated data access policy as infrastructure, not paperwork. Build accordingly.

Get the full brief at https://t.co/5Aqqb4isAs

If you want a personalized intelligence brief, check out: https://t.co/J5vykDVpoW

#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI

[mtowery @mtowery]

Welcome to Mike's AI Minute.

Today's big story is actually two stories that connect in a way most people are missing. The White House signed an executive order removing federal AI regulatory requirements. And Anthropic released Claude Sonnet 5, a frontier-adjacent agentic model priced to fit inside an enterprise margin structure.

Here's where people are getting it wrong. Federal deregulation doesn't reduce your compliance exposure. It moves the burden to you. Your clients still have internal governance boards. They still have EU AI Act obligations. Nothing changed for them. It just changed for you.

On the Sonnet 5 side, this is Anthropic's IPO narrative in product form. They're proving they can convert frontier research into recurring revenue. And for you, it means the cost justification for routing all your production traffic to Opus-class models just collapsed. The efficiency gap closed faster than most teams planned for.

My take? The teams that used regulatory ambiguity as a reason to delay building runtime guardrails and audit trails into their systems are now exposed. There's no federal floor to point to anymore. Build the governance layer now, while it's a choice. Not later, when it's a crisis.

Get the full architectural breakdown every weekday: 👉 https://t.co/5Aqqb4hUKU

Want a personalized intelligence brief? Check out: https://t.co/J5vykDURzo

#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI

[mtowery @mtowery]

Welcome to Mike's AI Minute.

June 30, 2026. Last day of the month. And it just became the most consequential month in AI history by almost every measure that matters.

Alphabet raised $84.75 billion for compute. That is not a typo. Noam Shazeer, the guy who co-invented the Transformer, left Google for OpenAI. John Jumper, Nobel Prize, AlphaFold, gone to Anthropic. Six senior researchers out of Google's AI org in five months. And GitHub Copilot's metered billing is hitting enterprise teams with cost overruns nobody budgeted for, because agentic sessions burn tokens at four to eight times the rate of standard autocomplete.

So here is what I want you to take from this. There are three things happening at once. Capital concentration, talent migration, and billing models that were not designed for the way we are actually using these tools now. Each one is manageable. All three at once, while your production workloads are live, is where teams get into trouble.

My take: the organizations that will have an advantage in Q4 are the ones making reversible architecture decisions this week, not the ones waiting for certainty. Pin your model versions. Cap your token budgets. Build the abstraction layer before you need it.

Read the full brief and subscribe for daily architectural intelligence: https://t.co/5Aqqb4hUKU

If you want a personalized intelligence brief built for your specific stack and role, check out: https://t.co/J5vykDURzo

#AIArchitecture #LLMOps #AgentSystems #EnterpriseAI #ProductionAI