Max aka Mosheh - AI with Apex

Most AI apps are stuck in a chat box.

Generative UI lets agents build tables, charts, and forms on the fly—while your app stays in control.

Use “UI intent” events (like AG-UI), render real components, send clicks/edits back as data.

Smarter models aren’t enough. You need smarter interfaces.

AI’s next edge isn’t “better chat.” It’s quiet automation on your own hardware.

Imagine texting a bot and coming back to:
• Inbox cleared
• Code deployed
• Local model installed to save money

The winners won’t just prompt well.
They’ll own the automations that run their business.

Hyperparameter tuning isn’t magic.
On a real student dataset, 4 tuned classifiers didn’t beat strong defaults.
With small, weak data + honest nested CV, performance barely moved.
Stop over-tuning.
Fix leakage, validate honestly, then invest in better data + features.

94% of CEOs are all‑in on AI.
The real shift isn’t a new app.
It’s your workflows quietly becoming AI‑assisted by default.

Ask: “What if AI did 80% of this and I edited the last 20%?”
The winners won’t be the best prompt writers.
They’ll be the ones who redesign how they work.

Your content problem isn’t “we don’t post enough.” It’s the chaotic workflow between draft and publish. 😵‍💫

Turn content into a system:
• Clear stages
• Owners for each step
• Automation for handoffs

Want a content engine, not heroic one‑offs? Fix the workflow in the middle. #ContentMarketing

94% of CEOs aren’t just playing with AI.
They’re doubling spend to up-skill people, automate routine work, and launch new products.

The real risk isn’t “AI takes your job.”
It’s someone who knows how to use AI doing it.

What’s one part of your work AI could help with, but you haven’t tested yet?

AI coding isn’t about faster typing.
It’s about a new role for you.

Imagine an IDE where you describe a game.
An AI plans, writes, fixes, and runs it.
You’re the architect. It’s the dev.

The winners won’t be the fastest coders.
They’ll be the clearest thinkers.

Mean imputation is like smashing a 3D puzzle and filling the gaps with beige Legos.

In my tests, mean/median beat KNN/MICE on accuracy.
But they quietly wrecked correlations.

There is no “best” imputer.
Only the best one for your goal:
Prediction vs. trustworthy relationships.

Researchers deleted 20% of a dataset on purpose.
Mean/median imputation beat fancy KNN & MICE on prediction.
But they quietly wrecked feature correlations.

Great for accuracy.
Disaster for clustering, PCA, and exploration.

Don’t just track error. Track what your preprocessing breaks.

Stanford’s SleepFM Clinical AI can read one night of sleep and forecast risk for 130+ conditions.

Trained on 500k+ hours of data, it uses brain waves, heart rhythm, breathing, and oxygen levels from a single overnight study.

The next edge in healthcare is earlier insight, not more data.

In 2025, AI stopped being a demo and became infrastructure.
PhD-level reasoning, agents running workflows, studios wired with gen-AI.
The tech isn’t the question anymore.
Your pace of adaptation is.
Have you moved AI from experiment to infrastructure yet?

Everyone’s hyped about “vibe coding” – describe an app, let AI build it.
AI is an accelerator, not a CTO.
Use it like a junior dev: keep humans on architecture, reviews, and tests.
Move fast, but still sleep at night.
What’s your experience shipping AI-written code?

Most people think AI risk is about superintelligence.
The quieter risk: we may never know if AI can actually feel.

A Cambridge philosopher says we have no reliable test for AI consciousness — and maybe never will.

Should we plan as if AI might feel pain, or act only on what we can prove?

“Open a tab” is dying.
“Give my browser a goal” is next.

Agentic AI browsers don’t just show the web.
They use it for you.
Read pages.
Click buttons.
Fill forms.
Plan trips.
Write reports.

Your job shifts from using the web to deciding what it should do for you.

A chip thinner than a human hair can now choreograph laser light like a symphony conductor.
Built in the same fabs as phone CPUs.
Less power, less heat, more control.

The real quantum edge won’t be more qubits.
It’ll be who can ship quantum like smartphones.

Agentic AI demos look magical.
Real deployments quietly fall apart.

The missing piece?
We judge agents only on final answers, not on how they plan, call tools, search, and use memory.

Log and train on the whole workflow.
That’s how you turn a demo into an adaptive system.

AI isn’t just for data centers anymore.
Tiny local models now run on laptops and Pis.
They chat, reason, see images, and stay private.
The next edge won’t be cloud-only.
It’ll be AI that lives right beside your data.
What’s stopping you from trying a local model this week?

Old-school audio editing is dying. Prompt-first editing is taking over.

Meta’s SAM Audio lets you:
• Type “remove barking dog”
• Click an object
• Mark a time span

Then it returns 2 tracks: your target + everything else.

The real skill now? Writing clear prompts, not knowing menus.

AI isn’t a black box.
Your words → tokens → vectors → attention → next-word bets.
Once you see it as structured focus at scale, you can redesign workflows, roles, and products around it.
What’s the first process you’d rebuild with this mental model?

AI didn’t fail in 2025.
Old workflows did.

Chatbots on top of bad data, clunky processes, and silos don’t create value.

AI only works when you redesign work end-to-end: clean data, modern infra, new roles, fewer handoffs.

Are you shipping bots—or fixing workflows?

Old-school “open another app” workflows are dying.
Adobe just moved Photoshop, Express, and Acrobat into ChatGPT.
You describe the edit.
AI does the heavy lifting.
The real advantage isn’t tools—it’s how fast your team moves when tools disappear into the chat.

You don’t need cloud GPUs to build with AI.

WebLLM, Free LLM Playground, BrowserAI, AgentLLM, and Genspark.ai let you run and compare models right in your browser.

No servers. No installs. No credit card.

What’s stopping you from trying browser-first AI this week?

Vibe coding is fun… until you realize you’re just copy‑pasting code you don’t understand.

AI can write code.
Your edge is thinking.

Shape messy requirements.
Understand data flows.
Debug production.
Design architectures.
Make tradeoffs.

The future belongs to devs who think, not just type.

Old school automation is dying. “Computer-use” AI is next.

Lux doesn’t just read the web. It clicks, types, and scrolls through real desktops at ~1s per action.

Think QA tests, forms, dashboards.
Trust is the bottleneck. Control is the unlock.

What would you hand off first?

Most AI reads more.
The next wave will remember better.

Apple’s CLaRa compresses documents up to 128x into “memory tokens,” then retrieves and reasons in that tiny space—often matching or beating full-text RAG.

The future edge isn’t bigger models.
It’s smarter memory.

Prompt engineering isn’t the real AI power move anymore. Context engineering is. It’s not about better sentences, it’s about better environments: profiles, memory, live data, history. The age of prompts is fading. The age of engineered AI worlds is here. Are you ready?

Prompt hacks are dying.

The real AI unlock isn’t “better prompts” – it’s better context.

Layered user profiles.
Live knowledge.
Selective memory.

Stop chasing magic spells.
Start designing the world your AI thinks inside.

An AI just scored 118/120 on the Putnam, but that’s not the real story.

It learned to write proofs, check its own logic, and fix gaps.

The edge won’t be who has AI, but who learns to reason *with* it.

Are you using AI as a calculator—or as a junior thinker with a verifier?

AI agent success isn’t uptime or speed. It’s goal accuracy, workflow adherence, and low hallucinations—tied to cost per outcome. Example: 62%→86% accuracy, 43% fewer hallucinations, $5.40→$3.10 per ticket. Use dashboards, audits, and daily tracking. What’s your take?

I tested Perplexity’s Comet. For static pages, it’s a rocket: fast summaries, cross-tab extraction, instant tables. On dynamic dashboards, it stalls. Use it as your research copilot, not a one-browser-fits-all. Want my 3-step playbook?