Michael Weed | Agentic AI Architect |

Salesforce, Google, & AWS Agentic Architect | HIPAA/PCI | Creator Atlas Engine + Atlas-G | Voice Coaching Anytime, Anywhere. Try free

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

AI safety is necessary. But when alignment becomes excessive refusal, degraded capability, or paternalistic product design, it creates a different kind of risk.

Users do not simply abandon the task. They migrate to local, open-weight, and overseas models that offer greater control, privacy, and reliability.

Qwen, DeepSeek, and other open models can already handle a growing share of work once reserved for frontier platforms. For many users, the gap is no longer large enough to justify higher costs, restricted outputs, and dependence on a cloud provider.

This matters beyond consumer preference. Regulated organizations are also exploring local deployment because it can provide stronger data control, auditability, and customization.

The West should not abandon safety but it must recognize that safety implemented without proportionality can weaken adoption, drive capability offshore, and accelerate the very loss of control it is intended to prevent.

We are not choosing between safe AI and dangerous AI but whether responsible, capable AI remains competitive. The answer looks grim in 2026.

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

Wheels up.

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

Friday closed one chapter: Agentforce AI Engineer supporting Jeppesen ForeFlight through Boeing and Beacon Hill.

Monday begins the next: Solution Architect, Agentforce and Data Cloud at Truffle Consulting, focused on Voice and Digital Engagement.

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

Aviation reinforced for me: AI augments expertise. Production value still depends on architecture, governance, domain knowledge, and execution.

I’ll continue writing about moving AI agents from PoC to production, especially across Agentforce, Data Cloud, Voice, and DE.

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

I canceled @grok Heavy because the shortness of the responses seemed to fit the narrative of soft sunsetting but I'm now on usage-based to see what they do.

I would think they would push for longer responses since I'm paying per token now but its still super short, sometimes one-word answers.

The experience feels broken and since we can run local offline models, this seems like a fast way to lose money for the business since I'm now looking at SillyTavern and other voice models with avatars.

Not sure the current business model but its de-prioritizing voice even as they realigned the voice architecture last in June for Grok Voice.

This makes me less interested in using Grok Voice and I would rather stick with Gemini Flash Live or Bedrock since they didn't de-prioritize the voice services I actually use. You don't want to rely on a service that changes with no transparency.

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

@rosawilts Gemini refuses completely normal business requests so this is not a surprise. 2026 is turning many people negative on sentiment of cloud frontier models.

2027 is the year more people start to discover local models, especially as new engines like Colibiri break the horizon

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

If xAI cuts Grok companions, this will be a huge blow. Some Heavy users pay $300 a month for more voice quota and this will lose paying customers + NPS will drop.

@AnimationInc made the companions so we might see new ones from https://t.co/XKHFA0gIZ7 or the Animates app

[Aaron @aaronp613]

SpaceXAI will be retiring Companions inside of the Grok app soon

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

@ShayWithCare @xai @grok It's the responsibility of the parents on what their kids can access. I always check in on what kids are doing and ask follow-up questions to know what they are doing or who they are talking to.

I don't want Zuckerberg or Musk being my children's parents and neither should you.

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

@xenovacom When you say "Kimi K3-like", what does this mean? Did you have to convert the base model?

Given the recent Colibri project, I wonder if Sol 5.6 or Fable 5 can optimize from a 1 token/sec proof-of-concept to something more usable in a similar optimization framework?

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

Today @SuperGrok disappoints once again.
1. Query multiple times the name of a basic meme but cannot find it
2. Found immediately with one @X query
3. Grok immediately recognizes it after sharing the meme

Why pay for Grok if X is better at surfacing info? How can you rely on it for anything with intelligent agentic orchestration when it fails at basic queries?

As much as I want to support multiple AI companies, X makes it difficult to recommend it to others for business.

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

The company that bans companies upon signup thinks it can do interplanetary communications? Is this written by AI or an intern?

You cannot even manage a social network; why would anyone think you can manage something actually important?

[Starlink @Starlink]

The Starlink team is exploring ways to extend connectivity beyond our planet

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

@Peeeeechy @Starlink You either are good at boosting, or you're an astronaut. You cannot do both

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

@MatthewBAX @Starlink No they wouldn't

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

@Starlink If you cannot fix X, you cannot do interplanetary communication. Please fix the one thats easy before thinking you are capable of the far more complex problem

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

@threenotes_jp Nothing because it does not matter. If someone calls you "weird", it means nothing.

If they say you are wrong, incorrect, or something else with value, you may consider it. Just don't let meaningless comments distract you

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

@esrtweet @cremieuxrecueil Plenty of code I've referenced online before AI has been just incorrect, and finding alternative working code has been challenging.

One thing I really like about AI is it comments a lot which people seem to hate, but I dislike inheriting code which has no documentation at all

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

@poetengineer__ This is a great 3D representation of relationships and thoughts. 2D lists being the default has perplexed me for some time as to why we have not progressed to something else.

Are you planning to expand it to anything else like browser bookmarks?

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

@Samaytwt Walk-in photography studios have almost disappeared from public life as a result of smartphone cameras.

Being "better" is not enough to keep a paid job. Devs moving forward need to offer less-friction, less pain, or something AI cannot.

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

This post, with nearly a quarter-million likes and 10M views, is made with AI

Do people not recognize this, or do they not care?

Its labeled "Made with AI", but if they don't care, they are willingly accepting AI-led narratives.

Interesting to consider the ramifications...

[Michael Weed | Agentic AI Architect | @MichaelWeedAI]

@fbjoe @billtheinvestor He didn't say but its for North Korea missile tracking only.

https://t.co/bOc9VPw0WM