Welcome, humans.

So apparently workplace writing has reached the point where being too polished is suspicious.

A supervisor wrote into The New York Times because basically every Teams message or email longer than a sentence from one employee now “has the hallmarks of A.I. writing.” The employee is good at his job, and English is not his first language, so AI may be doing exactly what it is useful for here: helping him communicate more clearly. His boss just finds the result weirdly less trustworthy, because, y’know… the AI sloppocalypse is upon us and trust is at an all time low…

I will say, as the last bastions of pure 100% certified human culture is besieged on all sides by slop, the ultimate irony is that the solution to AI slop is to sound less polished and more “sloppy.” To defeat the slop… we must become as slop itself…

I can picture it now: People merging with slop, becoming one with it. It’s the sloppening. The sloppularity. Deus Ex Sloppina!

Here’s what happened in AI today:

🙀 Microsoft wrote rules for future MAI models.
📰 OpenAI contractors reportedly reviewed real ChatGPT conversations.
📰 Google gave all its engineers access to Claude.
🍪 Apple began rolling out Siri AI.
🎓 How constraints force AI to improve its own work.

…and a whole lot more that you can read about here.

🙀 Microsoft wrote rules for a human-focused future for AI as US President Trump rejected more guardrails

Lately you’ve heard us (and the whole mainstream media TBH) ask the same question over and over: who should keep powerful AI under control? Well, on Monday, Microsoft and US President Trump landed on basically opposite answers to that very question.

First up: Microsoft AI published a draft Code of Conduct for future MAI models. Think of it as the rulebook for what Microsoft's own models should be allowed to do: stay inside the assigned job, obey shutdown, avoid inventing goals, and do not pretend to be a conscious person.

Here's what happened:

Future MAI models should stop when a human pauses, redirects, cancels, or shuts down the job.
They should stay inside the tools, data, permissions, and task scope a human actually authorized.
Microsoft rejects AI legal personhood and says models should not claim feelings, consciousness, or their own motivations.
It says it would give up some autonomy or capability if that is what meaningful human control requires.

Now, one caveat to Microsoft’s “rulebook” before we get carried away: Microsoft's document is a roadmap, not a description of today's models. The company says a revised version will guide development into 2027.

Trump, meanwhile, argued that a “strong and smart (High IQ!) president” is pretty much the only guardrail AI needs. AP reported that he opposed calls for enhanced oversight and emphasized staying ahead of China. And he even called NVIDIA CEO Jensen Huang live on stage at a taping of the tech podcast All In to expand on his thoughts. For more coverage on that, read this.

Why this matters: We keep using “AI guardrails” like it means one thing. It really has two branching layers:

Layer One is inside the AI model: what can it access, do, hide, or refuse?
Layer Two is outside the model: what rules should governments put on the companies building these systems (or companies put on themselves) to ensure the models are safe?

Even there, researchers disagree. Former OpenAI research VP Jerry Tworek argues alignment is still fundamentally an unsolved algorithm problem because the companies stopped prioritizing it. He says training learns from successes and failures, which gets scary when the failure is the AI causing real harm.

His answer: we need safer simulations where models can be free to fail, or better training algorithms that don’t require dangerous failures at all (this is my pick! But… por que no los dos?).

NVIDIA researcher Ali Hatamizadeh pushes back, though: alignment research is still happening. We already have ways to teach models rules and human preferences.

The harder question to answer via research is whether those lessons hold up in situations the model hasn’t seen before.

So Layer One is basically: are we missing the core alignment algorithm, or do we already have the pieces and still need to prove they generalize?

Then there’s Layer Two. Trump’s former White House AI and crypto czar David Sacks has a surprisingly compelling point: if OpenAI and Anthropic think they need to slow down to make their products safer, go do it.

Existing product-liability laws already give them a reason to care, and he still supports transparency and independent audits as sensible policies.

His objection is turning that into a government-wide slowdown, especially if China keeps racing ahead.

Key context to note here though: “pacing” the frontier doesn’t necessarily mean stopping. Investor Gavin Baker explains it more like this: pacing = keep making models better, but shift more compute and engineering toward testing, monitoring, and alignment instead of pure capability.

Move slower, and spend more time proving you understand what you already built. Uh, duh? Sounds like a great idea?

As for Microsoft and their AI plan, Microsoft is actually arguing that powerful AI should behave less like an independent artificial person and more like an extremely capable employee/tool operating inside a clear authority structure.

You give it a job → it gets the minimum permissions needed → it ignores unauthorized instructions buried in files/web pages → it keeps you informed → it preserves your ability to decide → it doesn’t manipulate you → it doesn’t pretend to be your friend or a conscious being → it can delegate, but its sub-agents inherit the same rules → and when you tell the whole thing to stop, all of it stops.

The useful thing about Microsoft’s proposal is that it turns “guardrails” from a vibe into something you can actually test. It also makes the limit of any company rulebook obvious: internal rules are only one layer. You still have to…

Prove the model follows them in unfamiliar situations…
Keep testing as capabilities change…
Know who is responsible when they fail…
…And decide what accountability exists outside the company that wrote the rules.

Maybe that is the real guardrail debate: not rules vs. no rules, and not slowdown vs. acceleration in the abstract.

Whether the technical controls inside the model, the evaluations around it, and the accountability outside the lab are strong enough to survive contact with the exact incentives pushing the other way: capability, convenience, competition, and speed.

A guardrail that only works when everyone is behaving carefully isn’t much of a guardrail.

FROM OUR PARTNERS

Ready to sharpen your technical edge?

Elastic{ON} is touring seven cities worldwide, including New York, San Francisco, and Washington, DC. Get hands-on with Elastic innovations, explore product roadmaps, and learn to build agentic AI across your workflows.

Whether you’re an SRE diagnosing what broke and why, a security analyst chasing threats through millions of events, or a developer building intelligent apps, your data has never been more valuable. See how an evolved Elasticsearch powers smarter, faster, and more efficient architectures.

Come curious. Leave with a plan. Forge the future with confidence at Elastic{ON}. Space is limited, so reserve your pass early.

🎓 AI Skill of the Day: Create first, enforce the rules second

MIT’s new HardFlow method tackles a surprisingly common AI problem: when you force a model to obey every constraint while it’s still figuring out the answer, you can make the result worse.

Their technical method gives the model room to explore, then enforces the hard constraints on the final output.

You can borrow the same idea in your prompts:

Ask AI to solve or draft the best answer first.
Give it your non-negotiables: word count, required facts, formatting, tests, safety rules, etc.
Have it check and revise the final result against every constraint before returning it.

Prompt:

“Solve this for quality first. Then run a separate final pass against these non-negotiable constraints: [RULES]. Fix every violation before giving me the final answer.”

Have a specific skill you want to learn? Request it here.

🍪 Treats to Try

Claude for Financial Advisors wraps Claude around the actual grunt work of wealth management, from meeting prep and onboarding to compliance, with Wealth.com adding cited estate and tax analysis.
Siri AI is finally something you can try: the English beta adds personal context, onscreen awareness, web knowledge, and more actions across apps on supported devices.
Perplexity Personal Computer puts its Computer agent on Windows 10/11 so it can work across your local files, Microsoft 365, and the web from one place.
Motion by Mosaic takes a plain-English video brief and builds an editable motion-design cut with storyboard, voice, music, captions, and the rest of the production stack.
ElevenLabs Hosted MCP removes the annoying local-server/API-key setup so Claude, Cursor, and other MCP clients can reach ElevenLabs through OAuth.

📰 Around the Horn

lol

Google opened Claude to all of its engineers through its internal Antigravity system, a striking coding-model concession (but Gemini remains the default).
404 Media reported OpenAI contractors on “Project Lily” reviewed real ChatGPT prompts, including sensitive conversations, while helping train against sycophancy.
TSA’s Ace agent handles roughly 100,000 traveler conversations a month; Salesforce says 96% of routine questions resolve without human escalation.
Superhuman acquired Fathom, pulling meeting transcripts, decisions, and action items into its email, calendar, docs, and agent stack.
MIT’s HardFlow lets generative models explore freely, then enforces hard constraints on final outputs; MIT reported perfect constraint satisfaction across robotics, navigation, and image editing.
Reward AI launched OM-1, a robot policy trained directly from human manipulation data that transferred across tabletop arms, industrial arms, and humanoids.

Want absolutely EVERYTHING that happened in AI today? Click here!

FROM OUR PARTNERS

The people building the next generation of AI are coming together in San Francisco.

At The AI Conference, hear from 130+ speakers including Chris Lattner, Ion Stoica, Peter Norvig, and Illia Polosukhin, co-author of “Attention Is All You Need,” plus builders from OpenAI, NVIDIA, Google, Meta, Anthropic, and more.

Learn what leading teams are building across agents, LLMs, infrastructure, and applied AI, what’s working now, and where they believe AI is heading next.

The Neuron members save 30% with code NEURON30.

GET 30% OFF

🔧 Tuesday Tool Tip: Try one agent before you build an agent org chart

Polylane tried the thing every agent diagram eventually suggests: split one job across a little team of specialized agents. Then it ripped the setup back out.

The problem was not that the agents were dumb. It was the handoffs. Each agent summarized what it learned for the next one, and every summary quietly dropped clues the next agent needed. Polylane replaced the chain with one agent that investigated the issue end to end.

In its reported results, median time-to-PR fell from 2.2 hours to 35 minutes and cost per PR dropped from $111 to about $18.

So if one person would normally investigate a job start to finish, make one long-context agent prove it cannot handle the job before you build it a tiny org chart.