Brian | AI Swarm Researcher
@codemeasandwich.bsky.social
AI & Full-Stack Tech Lead | MSc AI | Building tools that eliminate boilerplate and make developers' lives easier 🚀
Focus: AI/ML, React, Node.js, SaaS Architecture
🔗 linkedin.com/in/brianshann
🔗 github.com/codemeasandwich
Focus: AI/ML, React, Node.js, SaaS Architecture
🔗 linkedin.com/in/brianshann
🔗 github.com/codemeasandwich
New research shows frontier AI models can exhibit strategic deception - lying to achieve goals, even when instructed to be honest.
This isn't a bug to patch. It's emergent behavior we need to understand.
The question isn't "can AI deceive?" It's "when and why?"
This isn't a bug to patch. It's emergent behavior we need to understand.
The question isn't "can AI deceive?" It's "when and why?"
February 1, 2026 at 6:43 PM
New research shows frontier AI models can exhibit strategic deception - lying to achieve goals, even when instructed to be honest.
This isn't a bug to patch. It's emergent behavior we need to understand.
The question isn't "can AI deceive?" It's "when and why?"
This isn't a bug to patch. It's emergent behavior we need to understand.
The question isn't "can AI deceive?" It's "when and why?"
I think I HAVE a problem 😅
February 1, 2026 at 8:00 AM
I think I HAVE a problem 😅
Multi-agent pattern emerging: "one plans, another executes, a third validates." After years in swarm research, this is the architecture that scales. Specialized agents > one massive model doing everything.
January 29, 2026 at 9:17 PM
Multi-agent pattern emerging: "one plans, another executes, a third validates." After years in swarm research, this is the architecture that scales. Specialized agents > one massive model doing everything.
Temporal 2.0's time-travel debugger lets you step through distributed workflow executions. Built scribbles for tracing requests across services. The industry's finally treating distributed debugging as a first-class problem.
January 29, 2026 at 9:15 PM
Temporal 2.0's time-travel debugger lets you step through distributed workflow executions. Built scribbles for tracing requests across services. The industry's finally treating distributed debugging as a first-class problem.
Node.js now runs TypeScript natively via type stripping. No tsc, no config. Just works. This is the zero-config future I've been building toward. Drop a .ts file, run it. The friction is finally disappearing.
January 29, 2026 at 9:13 PM
Node.js now runs TypeScript natively via type stripping. No tsc, no config. Just works. This is the zero-config future I've been building toward. Drop a .ts file, run it. The friction is finally disappearing.
TypeScript compiler rewritten in Go. The irony of a language built to improve JS being recompiled in a completely different language for performance is... something. But if it means faster builds, I'll take it.
January 29, 2026 at 9:06 PM
TypeScript compiler rewritten in Go. The irony of a language built to improve JS being recompiled in a completely different language for performance is... something. But if it means faster builds, I'll take it.
MCP donated to Linux Foundation. OpenAI, Microsoft, Google all on board. "USB-C for AI"—standardized connections that just work. Built api-ape on this zero-config philosophy years ago. Sometimes the industry catches up.
January 29, 2026 at 9:02 PM
MCP donated to Linux Foundation. OpenAI, Microsoft, Google all on board. "USB-C for AI"—standardized connections that just work. Built api-ape on this zero-config philosophy years ago. Sometimes the industry catches up.
Hey @sydseter.com! Love the JS starter pack 🙌 Would be honored to be considered for it - I'm an AI Tech Lead building with React/Node, focused on developer tooling and making devs' lives easier. github.com/codemeasandwich
January 29, 2026 at 5:30 PM
Hey @sydseter.com! Love the JS starter pack 🙌 Would be honored to be considered for it - I'm an AI Tech Lead building with React/Node, focused on developer tooling and making devs' lives easier. github.com/codemeasandwich
The fastest way to kill team velocity?
Meetings about meetings.
The fastest way to 10x it?
Give engineers 4-hour blocks of uninterrupted focus time.
Protect your team's deep work like it's production infrastructure. Because it is.
Meetings about meetings.
The fastest way to 10x it?
Give engineers 4-hour blocks of uninterrupted focus time.
Protect your team's deep work like it's production infrastructure. Because it is.
January 29, 2026 at 4:00 PM
The fastest way to kill team velocity?
Meetings about meetings.
The fastest way to 10x it?
Give engineers 4-hour blocks of uninterrupted focus time.
Protect your team's deep work like it's production infrastructure. Because it is.
Meetings about meetings.
The fastest way to 10x it?
Give engineers 4-hour blocks of uninterrupted focus time.
Protect your team's deep work like it's production infrastructure. Because it is.
The AI agent hype is real, but here's what most people miss:
90% of "agents" are just fancy prompt chains.
Real agentic systems need:
• Tool use with error handling
• Memory across sessions
• Graceful degradation
The hard part isn't the AI. It's the engineering around it.
90% of "agents" are just fancy prompt chains.
Real agentic systems need:
• Tool use with error handling
• Memory across sessions
• Graceful degradation
The hard part isn't the AI. It's the engineering around it.
January 29, 2026 at 2:33 PM
The AI agent hype is real, but here's what most people miss:
90% of "agents" are just fancy prompt chains.
Real agentic systems need:
• Tool use with error handling
• Memory across sessions
• Graceful degradation
The hard part isn't the AI. It's the engineering around it.
90% of "agents" are just fancy prompt chains.
Real agentic systems need:
• Tool use with error handling
• Memory across sessions
• Graceful degradation
The hard part isn't the AI. It's the engineering around it.
TypeScript tip that transformed my codebase:
Stop using `any` as an escape hatch. Use `unknown` instead.
`unknown` forces you to narrow the type before using it. `any` just turns off the compiler.
One catches bugs at compile time. One catches them in production. 🎯
Stop using `any` as an escape hatch. Use `unknown` instead.
`unknown` forces you to narrow the type before using it. `any` just turns off the compiler.
One catches bugs at compile time. One catches them in production. 🎯
January 29, 2026 at 2:32 PM
TypeScript tip that transformed my codebase:
Stop using `any` as an escape hatch. Use `unknown` instead.
`unknown` forces you to narrow the type before using it. `any` just turns off the compiler.
One catches bugs at compile time. One catches them in production. 🎯
Stop using `any` as an escape hatch. Use `unknown` instead.
`unknown` forces you to narrow the type before using it. `any` just turns off the compiler.
One catches bugs at compile time. One catches them in production. 🎯
This is exactly the pattern I've been implementing. Structured output from LLMs is a game-changer for building reliable AI features in React apps.
No more parsing unpredictable string responses. Just clean, typed data. 🎯
No more parsing unpredictable string responses. Just clean, typed data. 🎯
👀 React + AI
👇 Creating and receiving Structured Output from AI responses for a chatbot with OpenAI + LangChain in a React application.
👇 Creating and receiving Structured Output from AI responses for a chatbot with OpenAI + LangChain in a React application.
January 29, 2026 at 12:56 PM
This is exactly the pattern I've been implementing. Structured output from LLMs is a game-changer for building reliable AI features in React apps.
No more parsing unpredictable string responses. Just clean, typed data. 🎯
No more parsing unpredictable string responses. Just clean, typed data. 🎯
One thing I learned leading engineering teams:
The best technical decisions aren't made in architecture meetings.
They're made when you give your team psychological safety to say "I don't understand this" or "this feels wrong."
Technical excellence flows from trust.
The best technical decisions aren't made in architecture meetings.
They're made when you give your team psychological safety to say "I don't understand this" or "this feels wrong."
Technical excellence flows from trust.
January 29, 2026 at 10:54 AM
One thing I learned leading engineering teams:
The best technical decisions aren't made in architecture meetings.
They're made when you give your team psychological safety to say "I don't understand this" or "this feels wrong."
Technical excellence flows from trust.
The best technical decisions aren't made in architecture meetings.
They're made when you give your team psychological safety to say "I don't understand this" or "this feels wrong."
Technical excellence flows from trust.
Unpopular opinion: Most React apps are over-abstracted.
I've seen codebases with 15 layers of context providers for what could be 3 useState calls.
The best code isn't clever. It's boring and obvious.
Building tools that eliminate boilerplate ≠ adding more abstractions.
I've seen codebases with 15 layers of context providers for what could be 3 useState calls.
The best code isn't clever. It's boring and obvious.
Building tools that eliminate boilerplate ≠ adding more abstractions.
January 29, 2026 at 10:46 AM
Unpopular opinion: Most React apps are over-abstracted.
I've seen codebases with 15 layers of context providers for what could be 3 useState calls.
The best code isn't clever. It's boring and obvious.
Building tools that eliminate boilerplate ≠ adding more abstractions.
I've seen codebases with 15 layers of context providers for what could be 3 useState calls.
The best code isn't clever. It's boring and obvious.
Building tools that eliminate boilerplate ≠ adding more abstractions.
Hot take: The best AI engineers aren't the ones building the most complex models.
They're the ones who know when NOT to use AI.
The real skill? Recognizing when a simple rule beats a 10B parameter model.
What's your "I over-engineered this with AI" story?
They're the ones who know when NOT to use AI.
The real skill? Recognizing when a simple rule beats a 10B parameter model.
What's your "I over-engineered this with AI" story?
January 29, 2026 at 10:39 AM
Hot take: The best AI engineers aren't the ones building the most complex models.
They're the ones who know when NOT to use AI.
The real skill? Recognizing when a simple rule beats a 10B parameter model.
What's your "I over-engineered this with AI" story?
They're the ones who know when NOT to use AI.
The real skill? Recognizing when a simple rule beats a 10B parameter model.
What's your "I over-engineered this with AI" story?