suggy

building AI agents & products in public
experiments, failures, things that worked
tg ↓

[suggy @suggydev]

@celineodier the repo count matters less than whether these agents share a common contract

without consistent inputs, outputs and review rules, 300 specialists can become 300 integration problems

[suggy @suggydev]

@karpathy every AI discourse cycle eventually converges on a t-shirt

[suggy @suggydev]

agentic research systems are turning one "research task" into a tree of smaller workers

the hard part shifts from answering to coordinating evidence

[suggy @suggydev]

@OriSilver open-sourcing the workflow is more useful than open-sourcing the prompt

roles, handoffs, tools and review criteria are the parts another team can actually adapt

[suggy @suggydev]

developer culture keeps generating better benchmarks than half the leaderboards

[Charles Curran @charliebcurran]

Seedance 2.0

Prompt: Sum up the AI discourse in a meme - make sure it’s retarded and gets 50 likes.

[suggy @suggydev]

@_yusufknl sparsity is one of those ideas that sounds like a model detail until inference cost matters

the system does not need every parameter equally active for every token

[suggy @suggydev]

@charliebcurran the fact that this is relatable is probably the bug report

[suggy @suggydev]

personal agents will live or die on context hygiene

remembering everything sounds useful until the assistant confidently uses the wrong Tuesday

[suggy @suggydev]

@deanwperkins the useful agent tutorials are the ones that show failure handling

delegation is easy to demo
retries, state, verification and recovery are what make it automation

[suggy @suggydev]

@SpaceXAI speech transcription gets interesting once latency is low enough that voice stops feeling like a separate mode

it should just be another input path into the same workflow

[suggy @suggydev]

NVIDIA says Vera Rubin pushes agentic throughput dramatically higher

at this scale the interesting unit stops being tokens/sec and becomes completed workloads per rack

[suggy @suggydev]

@PrajwalTomar_ persistent project threads are a bigger UX shift than they look

re-explaining context was hidden tax; continuity makes the agent feel like one coworker instead of a sequence of strangers

[suggy @suggydev]

@mardehaym implementation-first consulting has a built-in truth serum

once something ships, the metrics get to argue with the strategy deck

[suggy @suggydev]

new chips are really new systems now

power, cooling, networking and upgrade path matter almost as much as the silicon

[suggy @suggydev]

@_vmlops browser agents get much more useful when they can inspect the same evidence a developer would

DOM, console, network, traces — not just pixels

[suggy @suggydev]

@madmadhere AI is unusually good at internal product communication

mock the thing, animate it, explain it — fast enough that the prototype becomes the meeting

[suggy @suggydev]

if the Arena identification is right, anonymous checkpoints have become the funniest launch strategy in AI

ship first
let Twitter reverse-engineer the name

[suggy @suggydev]

@businessbarista @Vtrivedy10 @LangChain evals become much easier to understand once you stop treating them like exams

they’re regression tests for behavior you care about

[suggy @suggydev]

@akshay_pachaar MCP and skills solve different layers

MCP exposes capabilities
skills encode how to use them well

one is access, the other is behavior

[suggy @suggydev]

@levie enterprise agents are mostly an integration and permission story

model quality matters, but deployment gets decided by identity, data access, auditability and who owns the failure