This is either brilliant or scary:
Anthropic accidentally leaked the TS source code of Claude Code (which is closed source). Repos sharing the source are taken down with DMCA.
BUT this repo rewrote the code using Python, and so it violates no copyright & cannot be taken down!
Amazing how easy it is to hook up LLMs/chatbots today. I used Mistral's playground to create a custom Telegram chatbot for fun. Used botfather to create the bot. Whole process took about 3 hours end to end (research, code and crafting the prompt, deployment and secrets, test).
I send our monthly product update emails to the Reffie engineering team. Every time: “Wow, we actually shipped a lot this month.”
It’s easy to miss your impact when you’re deep in code. Step back a month, and it’s obvious.
We’re hiring: https://t.co/rA0O2L3no5
You can give GitHub Copilot custom review rules. Put them in .github/copilot-instructions.md and it will flag recurring annoyances like missing type hints or async issues. A small step that saves repeated PR comments.
Amid Meta's cuts to FAIR - calling all my friends at FAIR who are either impacted or know someone who was. I am hiring top-tier MLEs here in Toronto. Our mission is to make finding a place to rent suck less for everyone.
Top 2 signs you’re talking to a bad early-stage VC:
1. They won’t buy you coffee
2. Their value add sounds like everyone else’s but they think it’s unique.
If they can’t risk $5 to hear your pitch, they won’t risk $1M on your idea
Claude CLI v2 + Sonnet 4.5 isn’t just generating snippets anymore. It chained tasks, grabbed example files, and wrote me a working unit test suite while my tea steeped. I reviewed for a minute, no edits, pushed the PR. Big step forward.
https://t.co/MDwxev7QXF
Claude Code hooks let you run shell logic around AI actions. Block risky cmds (tf apply), auto-run tests/linters, or send async pings via Slack/Discord. Config lives in JSON per-project → deterministic, auditable AI. Docs: https://t.co/78jtsCXbhS
Protip: you can use memray to find strange memory uses. Take a look at our flame graph below. A function-level import for boto3 S3 (1 line fix) eliminates this memory use.
https://t.co/rjaWMVSYOk
Python 3.14 adds a new tail-call interpreter: enabled with --with-tail-call-interp, ~3% faster in benchmarks.
Strangely, it’s not the default. Why leave free performance on the table when semantics don’t change?
Docs: https://t.co/o5ieJh4JCE
#python#python3
Pro tip: memory leaks creep up in data pipelines, ML, and RAG workloads.
💡 Fixes:
• Run heavy jobs in separate processes (Celery, multiprocessing)
• Cycle workers with --max-requests in Gunicorn
• Import heavy libs only where needed
#python#ETL
Python 3.14 is adding template strings (t-strings).
Think f-strings, but safe for user-defined templates. Perfect for features like snippets, messages, and playbooks—where you want flexibility without exposing execution risks.
A quiet addition that will save engineers hours.
Poetry tip: poetry install --compile pre-generates .pyc files at build time instead of on first run.
In one Reffie microservice this means ~3,500 files precompiled, saving 200–500ms on container startup. Small, but it adds up across services.
Docs: https://t.co/B4sdZoXQKl
Multi-stage Docker builds cut our production image size by 75%. Build deps (compilers, bundlers, etc.) stay in one stage, artifacts get copied into a slim runtime image. Faster pushes/pulls + lower storage cost. Hard to go back after trying it. #docker#devops
I've been thinking a lot about how take-home interviews are broken.
We need to assume candidates will use AI. Stop testing “can they code.” Start testing “can they think.”
👉 Full post here: https://t.co/JFNwJHYPy9
Old take-home: build a Secret Santa app. Easy for Claude/Cursor—it’s already in the training set.
New take-home: assume AI scaffolds the code. What matters is the human layer—product vision, design choices, tradeoffs. That’s where skill shows.
Protip: You don’t always need ElasticSearch or OpenSearch for full-text search. Postgres has native support for text vectorization and ranking, which means you can often avoid the overhead of maintaining a separate search cluster.
#postgres#startup#programming
Client-side search is a deceptively hard problem.
Load too much data, and you risk overwhelming your user’s connection or device.
Pick the wrong library, and performance falls apart.
Best library I tested by far:
https://t.co/NIOJUfqnxB
#typescript#programming
Interesting paper making the rounds from OpenAI: https://t.co/Abi9dIeEIy
Choice quote:
Like students facing hard exam questions, large language models sometimes guess when
uncertain, producing plausible yet incorrect statements instead of admitting uncertainty.
#weeklyPaper