Germany
Joined July 2011
- Tweets2K
- Following649
- Followers710
- Likes2.9K
I feel like I developed quite an acuity for AI slop.
Maybe the same verbal fingerprint thing that anthropic is using to fingerprint AI text trained by reading tons of AI responses.
It's so easy to recognize AI writing now, even where people try to conceal it.
Not every product needs to integrate its own agent/harness. Some products are better off just implementing an interface (CLI or MCP).
I just want to connect my agent to your product, not be forced to use whatever you think is the best LLM / agent. Yours does not have my local memory.
Do I care too much about the wording of sales e-mails?
I find it extremely hard to use AI for sales, because of the amount of crappy AI slop copy I find in my inbox every day.
Does it even matter?
With open weights models trailing the frontier I always wondered if we wont at some point get to "good enough", where the frontier might be better, but the job can be done just as well with the last generation.
With Opus 4.8 I felt like the "unlock" for AI being usable for me happened, and anything newer is just slightly better.
Are there any good coding harness benchmarks out there?
I'm considering switching my harness, but I'd like to avoid sacrificing performance on task.
@thdxr Comparing RAM consumption among coding harnesses... is there maybe a real reason for needing so much RAM? (genuine question)
I read that @opencode uses *even more* RAM than Claude Code, and its RAM use already seemed absurd to me.
I.e. am I potentially missing a feature that obviously needs all that RAM, or do harness devs just don't assign much priority to it?
Do you remember the previous 100 chat messages or spoken words verbatim when chatting with someone?
Yeah, but your agent does, what a waste of tokens!
My agent's continuous context remembers only the past 2-3 turns and condenses into "short term memory" instead.
It trades off a lower cache-rate for less input tokens, faster inference/TFT, avoids context rot, no auto-compaction break, no session management overhead....
Just one long continuous conversation with the agent.
And btw, once any other context changes: persona files, memory pages, etc etc, your token cache is poisoned anyway, so now you pay for a huge 500k+ context.
It was my #1 cost problem with earlier versions of my personal agent harness.
While we're here.... @X using 400MB+ is similarly absurd.
At least you guys have to show images and videos, but I bet there's a nice 2x improvement hidden here.
@grok Do you think you could handle that for them?
Claude Code using hundreds of MB of RAM is still absurd to me.
@ClaudeDevs Any chance you could run a prompt to optimize that pretty please? My naive C/C++ dev brain tells me 3x improvement might be low hanging.
I think long-haul agentic coding projects are still very undersaturated.
Work on something for long enough, it far exceeds any "one-shot prompt demo slop", compound the dev speed into something crazy.
I've just seen glimpses of it so far.
Claude Code using hundreds of MB of RAM is still absurd to me.
@ClaudeDevs Any chance you could run a prompt to optimize that pretty please? My naive C/C++ dev brain tells me 3x improvement might be low hanging.
And the winner is...
... @CocaCola.
Does Samsung Photo app on Android really force Samsung Pass when scanning a passkey QR code instead of respecting the passkey settings?
Bad call.
I write my personal AI assistant harness myself.
OpenClaw, Hermes, etc etc are all really nice, but I enjoy knowing every corner and risk of the agent's harness and questioning the status quo.
For example, my agent doesn't have "sessions" like others, it's one continuous conversation with "working memory". Memory is a hierarchical tree, approvals/permission are treated slightly exotically, and so forth.
Love it. Zero context opinion drops doing so well on X show how well-optimized the platform increases regrettable minutes spent.
Unfortunately ngmi.
Anyone else find themselves repeatedly need more powerful CI runners for Rust or C/C++ compilation?
cirunner.dev/
Hey Rust/C/C++/wasm/etc devs with long build times:
Is this interesting to anyone else than just me?
cirunner.dev/
Opus 5.5 rewrote PowerPoint in a single prompt...
... is probably what I should have written for X, but it took many hundreds of prompts, over months. And it was mostly Opus 4.8.
Full PPTX IO and rendering engine in ~320kb #WASM (C/C++), fully client side, renders in milliseconds with @WebGPU, with PPTX, PDF and ODP export.
Supports ~76% of PPTX features on import.
UI is far behind the rendering features, desktop only for now.
👉webdecks.app/app/?open=DemoD…
