family mooncake 😋

He will give practical advice, and concrete criteria to consider, when choosing research projects, and making professional decisions, in these last few years before AGI." docs.google.com/presentation...

🚀🚀🚀

You already know the answer, but it’s nice that someone put in the manual effort to create a benchmark Can AI file your taxes? No "TaxCalcBench: Evaluating Frontier Models on the Tax Calculation Task" Paper: www.arxiv.org/abs/2507.16126 Repo: github.com/column-tax/t...

Job update: a couple of weeks ago, I joined @tensorlake.ai full time. I’m having a lot of fun building the product with @diptanu.bsky.social and the rest of this wonderful team. We have a few open positions if you’d like to work with us: www.linkedin.com/jobs/search/...

On to new things! "Ingest, query, and share telemetry data with your engineers and customers at a fraction of the cost."

Regarding goroutine & unbuffered channel interaction, found myself repeating this multiple times. Maybe it's time to write it down for reference. Early return + unbuffered send = goroutine leak. rednafi.com/go/early_ret... #golang

1/ SlateDB v0.8 is now available! This is release includes OpenDAL object store support, serializable snapshot isolation, first-class Go bindings, Python binding improvements, deterministic simulation tests, performance improvements, and tons of bug fixes. Details below. 👇

2/ Snapshot isolation — @flaneur2024.bsky.social has been hard at work on snapshots/transactions. 0.8 now has `DbSnapshots`, which provide a consistent point-in-time DB view. Sequence numbers are now core to SlateDB and will be used for many features (including transactions) going forward.

it seems there's no gold-standard benchmark for coding models at the moment. imho, the only way to know if one is any good is to throw a (real world) task at it and see if it can handle it smoothly 🤔.

OpenDAL can now use object_store as its backend service 🥳! this allows users to leverage the out-of-box primitives from the OpenDAL operator, such as chunking / parallel fetching and caching on user-provided object_store instances github.com/apache/opend...

in the age of vibe coding, the measure of a top programmer should be how long they can get Claude Code to work continuously without human intervention. if you can get Claude to code independently for a full 24-hour workday, you deserve top dollar in the market 🤔

z.ai's slime It is built with native SGLang integration, carrying its full inference optimizations straight into training, is an LLM post-training framework for RL scaling, providing two core capabilities:

SlateDB now has snapshot support! so the next step will soon be (possibly SSI?) transactions 😁 github.com/slatedb/slat...

I went through DuckDB's WAL, and it does everything I was asking for in my blog post: 1. Per record checksum 2. Explicit error on checksum failure 3. Configurable behavior 4. Partial recovery 5. Safe truncation of the WAL only when WAL contents are checkpointed

yesterday a friend told me a narrative about why ResNet works so well， he said that some patterns can already be learned by a shallow, 3~4 layer FFN. when we naively stack more layers, the extra layers are forced to learn an `identity` function 🤔

recently Kimi K2 has become my most-used LLM for chatting. imo, the best part is that it does not think 😮. while thinking is useful to improve the final result, fast fact retrieval is often more valueable for me than more 'correct' answers.

made a Homebrew formula for FizzBee 😁, now we can install fizzbee with: brew tap fizzbee-io/fizzbee brew install fizzbee github.com/fizzbee-io/h...

📔Writing our Rust database from scratch gave us a golden opportunity to implement deterministic simulation testing as a core feature. Learn about the architecture that gives us full control over DST ingredients. 🧊 www.polarsignals.com/blog/posts/2...

Whew. This one was some work. It's still very simple, but already has found some interesting stuff.

i found what vibecode really does for me is help me jump in the first line of code earlier instead of overthinking everything.. i heard INTJ people tend to have trouble taking action until they've got everything perfectly planned out..

i have a bias that the more powerful a programmer is, the more fully they recognize the capabilities and potential of AI programming 🤔 antirez.com/news/154

this is obviously the use case where io_uring best showcase its capabilities 😲 tangled.sh/@rockorager....

So.. Pierre Barre just built an NFS server implementation on SlateDB. ZeroFS is today days old, but it's going to get wild. Bottomless storage and checkpoints mean instant FS snapshots (zero overhead). github.com/Barre/zerofs

excited to have finally gotten this RFC merged in SlateDB 😆. the implementation around this has already moved a lot github.com/slatedb/slat...

They research "What Makes a Base Language Model Suitable for RL?" or why RL on LLaMA consistently underperforms Qwen. Their key insights: - High-quality math data is key to RL scaling. - QA-style data helps, but it depends on task similarity. - Instruction data boosts QA’s effectiveness.

Insane amount of SlateDB work going on: - snapshot reads - split/merge DBs (zero copy) - deterministic simulation testing And someone just pushed Python bindings in a PR! 🤯

the essence of vibe coding seems to be about being a good rubber duck for LLMs. it's essentially just another form of rubber duck debugging 🤔.

just noticed that we can't simply use std::sync::Mutex within a tokio task when there's some .await points living in its guard, because MutexGuard is marked as !Send, but the tokio tasks are possible to get scheduled across different threads 😮... we must use tokio::sync::Mutex at this situation..

Responsive talkin about their storage engine for Kafka streams. (Built on SlateDB!)

How Hugging Face migrated terabytes of model data from Git LFS to their own storage backend, called Xet: huggingface.co/blog/xet-on-...

i'd heard Volcano is widely used for running AI workloads on k8s. my initial quick look only showed me that it added batch processing support to k8s. batch processing - simply a task queue to run tasks one by one, why bother such complexities?

0x.tools xCapture v3: Linux Performance Analysis with Modern eBPF and DuckDB 👀👀 tanelpoder.com/posts/xcaptu...

٩( 'ω' )و

opinion: a physical printer is the best read it later app 🤔

Highly recommend this!

We need more of this in the discourse: AI is normal technology. I actually disagree with the worldview articulated here, but it's closer to the truth than the superintelligence/singularity view IMHO.

this weekend's paper reading is 'Sarathi: Efficient LLM Inference by Piggybacking Decodes with Chunked Prefills', which discusses Chunked Prefills.

I have some thoughts about boringtechnology.club for you, as the work turns ten years old this weekend.

Microsoft's Playwright, that enables reliable end-to-end testing for modern web apps ( playwright.dev ) built a Playwright MCP ( github.com/microsoft/pl... ) that provides browser automation using Playwright. This enables LLMs to interact with web pages through structured accessibility snapshots.

Fastrace provides a production-ready solution with seamless ecosystem integration, out-of-box OpenTelemetry support, and a more straightforward API that works naturally with the existing logging infrastructure. fast.github.io/blog/fastrac...

WASI p3 is almost here! dev.to/fermyon/look...

realized the difference of MCP server and the Function Calling in ChatGPT is that MCP is an protocol that interacts with the local clients, it communicates with the client with stdio. so it makes it possible to interact with the softwares installed on the desktop like Blender 3D.

recently i learned a trick about using LLM is: let LLM to write a better prompt for me, then use the generated prompt for work. 🤔 the generated prompt often has a more comprehensive consideration for the task 🤔

After nearly more than 5 months of discussion, SlateDB has accepted an RFC for synchronous commits and durability. This will be the foundation for transactions and snapshot serializable isolation. Thanks to @flaneur2024.bsky.social for powering through this! github.com/slatedb/slat...

read sqlancer's paper and found its idea is pretty simple & effective 😲: it picks a random row as a "pivot row," then generates a BOOLEAN condition with this row's random columns that returns TRUE, and validate whether the pivot row is included in the result or not.

i found the experience of using Cursor with Gemini Flash 2 is pretty nice for it's simply pretty fast response. 😲 let me pause the subscription of Cursor for a while..

Talk 24: "Serverless primitives for the shared log architecture" by Stephen Balogh (@sbalogh.bsky.social)