Pinned Tweet
finally added GitHub Sponsors and ngl I’m so happy about it lol
I keep building weird open source stuff around agents, infra, browsers, debugging, reliability and whatever else looks fun enough to break
if any of it has been useful to you, #sponsor me so I can spend more time building even weirder stuff
github.com/sponsors/ixchio
Aman retweeted
DolphinBench: Mapping the Pareto Frontier of Agent Memory [dolphinbench.ai]
Opus 5.5, one-shot, no instructions, 20mins (sound ON)
Today we are releasing DolphinBench: Mapping the Pareto frontier of agent memory.
Coding, tool use, and long-context retrieval already have strong public benchmarks. Memory still mostly gets graded as a quiz: ask a question, judge the answer, maybe report precision and recall.
A lot of those boards are getting saturated. They also skip the two questions that matter when you pick a memory system to ship:
At what cost?
At what latency?
Agents are not taking memory quizzes but they take actions. The fact that matters is usually missing from the request. Either it shows up in the tool call, or the agent does the wrong thing while sounding fluent.
DolphinBench grades that action after long history, and puts accuracy, cost, and latency on one board.
dolphinbench.ai
Cloudflare’s Sep 22 update is a big MCP infra signal:
Private MCP servers can now sit behind Cloudflare Gateway instead of being exposed publicly.
that means agents can access internal tools via identity-aware routing:
agent → MCP portal → policy → private route → tool server → audit
MCP skill alone won’t stand out but the secure MCP infrastructure will.
Cloudflare just made Python Workers GA
the interesting part isn’t “Python on the edge”
it’s that FastAPI, Flask, Django, LangChain and MCP can now run inside Workers through Pyodide + WASM
that means you can build a Python AI backend with
Durable Objects for state
Queues for async work
R2 for files
Vectorize for retrieval
Workers AI or external models
without owning the usual app server stack
Python + WASM is starting to look way less experimental
Aman retweeted
Inspired by Jev, we built Open-Jev: open-source decision models.
2B/9B: LoRA adapters + decision heads.
Code: github.com/Zefan-Cai/Open-Je…
Data: huggingface.co/datasets/Zefa…
2B: huggingface.co/ZefanCai/Open…
9B: huggingface.co/ZefanCai/Open…
Watch the demo ↓
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
fuck Jev, use Laya
not every decision needs an LLM generating text back at you
Laya takes state + typed questions and returns calibrated choices, scores and yes/no probabilities in one forward pass
runs from Node + TypeScript through ONNX
no Python or PyTorch at runtim ~140ms for 3 batched questions on Apple silicon
this is the kind of weird small model infra I want more of
github.com/receptron/laya
huggingface.co/convaiinnovat…
How I built Laya
laya.convaiinnovations.com/
RubyLLM 2.0 is kinda wild
tool approvals, citations, caching, batches, OCR, reranking, speech, video, durable agents and Rails persistence all behind one Ruby API
Ruby is quietly becoming a legit stack for shipping AI apps
RubyLLM 2.0 is the biggest release of my career.
Don't believe me?
Check out the release notes.
github.com/crmne/ruby_llm/re…
starting to think round robin is a pretty dumb abstraction for agent traffic
an agent request is not really a request, it is a session that disappears into a tool call then comes back with the same context 400ms later
the scheduler that remembers that is probably going to beat the one that only sees queue depth
anthropic says claude is already doing 26% of its own model R&D work under human supervision
the interesting part isn't self improving AI yet, it's that the bottleneck starts moving from writing the code to deciding what work the model is allowed to do, how you verify it, and when you stop it
this is kinda funny but for agent inference the least loaded gpu can actually be the wrong gpu
if the next turn lands on the worker that already has its kv cache sitting there keeping the session there can beat moving it to a quieter gpu
vllm tested this on AgentX and simple session aware routing beat balancing by queue depth or kv usage
we spent years teaching distributed systems to spread load and now agents are making locality matter again.
every local model demo looks impressive until you ask what happens after 30 minutes of real use
memory keeps climbing context gets expensive latency starts drifting
that's the benchmark i care about
agent failures are rarely “the model is dumb”
usually it’s some boring shit in the runtime
bad state duplicate tool calls no replay no timeout policy zero observability
fix the system before swapping the model
the shift : what huge developer behavior just changed, and what primitive is still annoyingly missing?
anthropic dropped mhs yesterday
model hardware standard
ai agents can now operate physical devices
microscopes liquid handlers robotic arms
lab and factory equipment
this is the digital to physical bridge everyone talked about now its actually happening
not chatbots controlling screens
actual agents manipulating actual hardware