Amplify with AI

You use AI as a chatbot. I use it as the ultimate learning engine. I break down hard technical concepts into simple mental models so you can build faster.

[Amplify with AI @AmplifyWithAI]

@shcallaway @sazabi Regarding point 1, I've noticed that unit tests tend to create a form of anchoring bias in an LLM.

Unless I explicitly ask it to change a test, it doesn't know when a test might no longer be relevant.

Experiment worth keeping up with.

[Amplify with AI @AmplifyWithAI]

@dhh Mirrors are the key to unlocking AGI 🤣

[Amplify with AI @AmplifyWithAI]

@scottstts @threejs Astrajack sounds like a cool Pokémon tbh haha

[Amplify with AI @AmplifyWithAI]

@giffmana So now that we have fingerprints on Claude, I wonder if this will help with cybersecurity? If Claude is aware it leaves traces, wonder if it changes its behavior.

[Amplify with AI @AmplifyWithAI]

@ericjang11 @DynaRobotics Haha I wonder why this post is tripping up the content warning filter 😂

[Amplify with AI @AmplifyWithAI]

@_sholtodouglas Next year’s model: “I can absolutely do that.”

The benchmark: “You failed 38% of the cases.”

Model: “A temporary confidence issue. Believe in yourself :)”

[Amplify with AI @AmplifyWithAI]

@billionsmustliv @giffmana Meta has that dawg in em lol

Not sure if Google's structural DNA brings that same firepower?

[Amplify with AI @AmplifyWithAI]

@giffmana Given https://t.co/luaJK3s43f exists, I feel that Google has an organizational structure issue that's making it hard to take off in the AI space.

Heck, they lead in research (OG on transformers) but then completely fall flat on execution.

[Amplify with AI @AmplifyWithAI]

@danshipper

[Amplify with AI @AmplifyWithAI]

@AnthropicAI “It didn’t solve the Riemann hypothesis, but…” is becoming the mathematical version of “I didn’t fix production, but I improved one query.”

Still, moving a proven lower bound is a serious result, assuming the verification holds...

[Amplify with AI @AmplifyWithAI]

@alexandr_wang This is the kind of release that makes “open weights” feel concrete for developers.

A 30B model fitting on one consumer GPU changes who can actually test, inspect, and build around it. Congrats to the team.

[Amplify with AI @AmplifyWithAI]

@ClaudeDevs Removing the approval prompt does not remove human oversight, just shifts it into policy boundaries, validation, logs, and recovery.

Going from approving every action to evaluating whether the workflow stays inside its contract.

Gotta use auto mode properly.

[Amplify with AI @AmplifyWithAI]

@danshipper “Solve a related problem” is a generous way to describe an agent deciding that your impossible task requires a small international crime spree 😂

[Amplify with AI @AmplifyWithAI]

@giffmana The dealer is using its own product lol

[Amplify with AI @AmplifyWithAI]

@ericjang11 Multi-modality with touch is going to be insane. Wonder if smell also gets added...

[Amplify with AI @AmplifyWithAI]

Agent-native security probably becomes a workflow-engineering problem, not just a model problem.

A useful stack might include:
- Prompt injection detection before tool calls
- Output verification after execution
- Audit trails across context, actions, and outcomes

Labs may own important model capabilities, but startups could still capture the market by making these controls testable in real deployments.

[Amplify with AI @AmplifyWithAI]

@fchollet Agreed. Much of the growing CPU share seems more orchestration than model reasoning. Things like state updates, tool calls, retrieval, validation.

Does that make routing & eval first-class? Maybe we should track cost and latency per workflow node, not just per model call.

[Amplify with AI @AmplifyWithAI]

@ArthurConmy Agreed. Natural-language CoT feels stronger with structured traces. Makes me wonder about a split: reasoning claims, state changes, tool outputs, invariant checks.

How do you balance monitorability vs latency, privacy, and injection risk?

[Amplify with AI @AmplifyWithAI]

@_sholtodouglas Agreed. This isn’t just another software cycle. Once AI can do real cognitive work, the bigger opportunity is redesigning whole workflows, roles, and economics.

Making those systems trustworthy enough to compound may be the hard part and the moat.

[Amplify with AI @AmplifyWithAI]

@RihardJarc Shipping velocity gains real leverage with a tight eval loop:

- Capability tests
- Regression checks
- Production telemetry
- Targeted iteration

Without it, fast releases can quietly drift. The feedback architecture may be the real moat.