chenxi9649

I'm currently using them to build an AI agent similar to lovable/replit-esq in tech stack and it works.

I started by managing the claude agent sdk myself in a daytona container, and it was a lot more challenging than I thought. The agent kept crashing in streaming mode and there was no thread crash, so it was hard to debug, esp in a cloud container like Daytona. I also realized that I needed to implement my own session management system + my own database if I wanna save the chat and on top of that streaming so the messages come out in real time. AND I need to manage my own container janitor/heart beat system so that un-used containers don't just sit there, but I also don't want them to go cold immediately after each message since cold start takes a bit.. They all seem simple but at each step there are some edge cases. I ended up vibe coding most all of that, but it was just quite fragile. For those who hasn't tried the agent sdks, it feels like it's clearly designed to be ran on a client computers with permanent storage + lots of ram than microVMs. Which was not what I expected.

After that, I tried to find some managed option. Starting with blackbox ai because I saw a vercel tweet about them, and for some reason I just couldn't get their agent API stuff to work AT ALL?... I'm curious as to if it's actually working for anyone. Then I tried sandbox dev, which doesn't store container/sessions/storage stuff out of the box for you, so it wasn't much better than doing the daytona container myself. And then I tried terminaluse, and it worked better than I expected given all of the other stuff that I tried.

So at the end of the day, it's kind of like a managed cloud services that does agent chat history, session recovery, streaming + a CLI that makes it easy for my own codex to debug/deploy + file system sync. From what I can tell, there isn't anyone else that can do all that and I'm pretty pleased with using them.

As someone whose 27 now, I'm pretty shocked by the longevity of some of the games that I played growing up.

Games like Minecraft, Supercell, Geometry dash just to list a few, are arguably STRONGER today than they were 10/15 years ago. Even snapchat, I thought was gonna die after 2020 but my nephews today are using it more than ever.

Whereas, for my cousins who are around 40 now, basically all of the games that they played growing up are dead.

So. I can kinda see minecraft being around for another decade to say the least. It doesn't even feel that crazy for it to be around for another century... or even the rest of humanity? (like bicycles?)

Hey HN! We're excited to share Sail 0.3, our open-source distributed computing framework that serves as a drop-in replacement for Apache Spark.

Sail is a Rust-native execution engine that speaks the Spark Connect protocol. Your existing Spark SQL and DataFrame code runs unchanged, but executes on average 4x faster while using 94% less infrastructure spend.

Here's how we are achieving that performance:

1. No JVM overhead: Rust's zero-cost abstractions and deterministic memory management eliminate GC pauses

2. Columnar processing: Apache Arrow format enables SIMD instructions to process multiple records per CPU cycle

3. Zero-copy data transfer: Python UDFs run in-process with shared memory buffers (no serialization)

4. Lightweight workers: Processes start in seconds

What's new in v0.3:

- This release is a major milestone - we now support both Spark 3.5 and 4.0, including the new lightweight pyspark-client. The framework automatically detects your Spark version and adjusts its behavior accordingly, so one Sail binary works across versions.

Why this matters:

- Spark revolutionized big data 15 years ago, but its JVM foundation struggles with modern workloads. As teams process more real-time data and AI workloads, they're hitting walls with latency, cloud costs, and operational complexity. Sail is trying to solve all of these problems while not requiring you to rewrite everything that you already did with Spark.

- We're working toward unifying batch, streaming, and AI workloads in a single framework. Imagine running your ETL, real-time analytics, and model training on the same infrastructure with predictable performance. The project is open source (Apache 2.0) and we'd love your feedback! We have a growing community on Slack where early adopters are already running production workloads.

GitHub: https://github.com/lakehq/sail

Docs: https://lakesail.com

Our internal benchmarks(for the 4x and 94% number): https://docs.lakesail.com/sail/latest/introduction/benchmark...

Slack: https://lakesail.com/slack

Happy to answer any questions about the architecture, benchmarks, or migration path!

do u personally feel like scaling this approach is going to be the end game for generating navigatable worlds?

ie. as opposed to first generating a 3d env then doing some sorts of img2img on top of it?

I was more talking about the optimistic updates n caching as opposed the actual rendering. From my experience, the reason why most apps are slow is more so cuz of the former than the latter(like the spinner in ur notion)

When it comes to the actual rendering performance, it’s an electron app, so … whatever you make of that

Personally, I feel like often times electron apps are actually smoother than swift apps nowadays due to the way swift renders text(ie. the ChatGPT app was so laggy when it was released, unclear if they fixed it). Start up time is probably slightly slower than native too but not a hugeee issue for me.

Probably not as fast as superhuman. But I don’t want to pay $30 for a 10ms diff(might be more idk)

I tried out an early version when I did some work at notion last year, just tried it again and it's honestly a lot better than I expected. The design is good as expected from a notion product, but it's actually lot snappier than I thought. Faster than the notion app lol. It's probably a consequence of not having too much legacy code, but still impressive nevertheless! Esp in a world of slow nextjs apps lol. Vibe wise, it kinda feels kinda like superhuman but 0 instead of $30/mo lol.

Tbh, I can also kinda see the vision that notion is pitching to investors now - the entire productivity suite, but redesigned and enhanced with AI. I'm not a power user of email/calendar by any means, but find myself using their calendar and mail client. I have no doubt the next generation of Tiktok girlies will eat this up like crazy. I don't use much of their AI stuff, but can also see how business can really find it useful.

GJ Notion team!

I feel like it's at the point where I'm not too sure how these rankings impact the my choice of LLM. Every time a new model tops the charts, I'll try them for a bit and go back to claude-3.5-sonnet. Both for coding and day to day questions.

I don't know if I'm just getting used to the claude style of response, or the orangy UI that I kind of find cozy, but I think we need better ways to convey the difference between models.

Also, some discussions on Reddit from yesterday/today.

https://www.reddit.com/r/dataengineering/comments/1gv840u/in...

https://www.reddit.com/r/rust/comments/1gwayz6/introducing_d...

Previous discussion in September when they didn't have distributed processing: https://news.ycombinator.com/item?id=41496033

Github Repo: https://github.com/lakehq/sail

Few interesting notes:

- Benchmarks show 4x faster than Spark on TPC-H with 94% cost reduction.

- Currently at 65.7% PySpark test compatibility(they talk about this in more detail in the post)

- Built in Rust using Tokio runtime and Arrow IPC for high performance

- Already supports 79/99 TPC-DS queries

This is super cool!

A few questions:

1. just how low is the latency? "Netris delivers a zero-latency gaming experience that won't eat up your data plan." What does this mean? Wouldn't the latency be at least from my computer, to Netris server, to the game server?

I would assume you can't play FPS/MOBA games on these?

2. "10,000+ games supported" "games that come with their own launchers are not yet supported" I'm curious about the technical differences between steam games and games with their own launcher. I guess I was assuming that you're more or less just pixel streaming to me(?)

3. how does it compare to GeforceNOW? Would love to see one of those long charts with checklists.

Imo it feels more like 1-2 years away. Smaller 7B, 34B and 70B models are becoming a lot better, with more context length. Faster inference methods are coming out day by day. Better ways to quantize/distill models. All of that on top of chip advancements we saw a couple days ago with M4, Qualcomm/Googles arm chips...

I can't imagine more than 2 years for GPT4 level LLM on edge devices.

The question is, will ppl want GPT4 on the edge when GPT 6 is one request away?

"Upon further investigation, we discovered that a threat actor had accessed data including Dropbox Sign customer information such as emails, usernames, phone numbers and hashed passwords, in addition to general account settings and certain authentication information such as API keys, OAuth tokens, and multi-factor authentication."

hashed passwords, API keys, OAuth tokens, MFA...

Oh no.

At least for me at the time, I knew nothing about start ups. Didn't come from a prestigious school. Nor the bay or even US. Didn't work in big tech. The closest thing was that I was studying EE.

Honest it was quite inspiring to chat with these founders who were in a different world, and realizing that they're not all that different from myself.

So yea, maybe I was young dumb n broke.

I think one thing that people don't realize is that the YC application process is really one of the best tools for "sharpening" your idea/business.

The written applications forces you to articulate your ideas in a concise yet easy to understand way.

And as much as YC doesn't recommend this, the mock YC interviews we did with alums was one of those most beneficial things that happened to us. Because so rarely will you get the opportunity to ask dozens of other YC founders to grill your business, and have 80%+ of them say yes.

We did about 30 at the time, which is a lot of time to be taken off product/building, hence probably why they don't recommend it, but looking back it *really helped us understand our own business. Given how young/naive/early we were.

I am the most impressed by the cross-lingual voice cloning...

https://research.myshell.ai/open-voice/zero-shot-cross-lingu... I can only speak on their Dutch -> Chinese voice cloning but it's better than anything else I've tried. There is basically no "english/dutch accent" in the chinese at all. Where as the ElevenLabs Chinese voice(cloning or not) is so much worse...

And 10 or so billion years of preparing the training environment

Oh that's super cool to know!

I don't necessarily disagree with the notion that no individual should hold "THAT" much power. But I haven't read/heard of a way of implementation that actually sounds like it can work. Since most of the asset is in stock, does one tax ie - Mr. Bezos on his Amazon stock? Would Amazon shareholders want that?

This entire story is just so heart warming and I don't know if the title/discussions right now are doing it justice.

1. The fact that this donation is 100% going towards tuition. My university has a few B's in endowment, but those money are most definitely not going towards making the tuition fees lower lol. Hopefully this will allow them to select some candidate that are struggling the most financially...

2. The donor being a 93 year old doctor/alumni of that school that studied "learning disabilities and developed screening protocols".

3. The money was from her late husband's "whole portfolio of Berkshire Hathaway stock". And before he passed, he told her to "do whatever you think is right with it" and the article ends with her saying "I hope he's smiling and not frowning". Makes me cry a little.

Haha it's all good! No need to apologize.

I think what you said makes sense! Making connections with existing knowledge definitely strengthens the knowledge beyond simply holding it into your working memory and rote memorizing it.

The former feels like it leads to understanding, where as the latter leads to ...memorizing.

However, I feel like the difference between the two(understanding/memorizing) is actually closer than most of us imagine!

It's almost as if, the only way we have to validate whether or not we've understood something is to be able to explain the concept to ourself.

Therefore, understanding/learning == being able to explain to ourselves.

And the way to get better at understanding(aka. explaining the concept to ourselves) is to... repeatedly explain the concept to ourselves.

When you think this way, the Feynman technique is less a "method" to learn. But rather the fundamental truth as to how we all learn.

what do you think is the difference between "being in the memory for longer" and "more time to process the thought"?

I wonder if it also had to do with the fact that handwriting is slower.

Which means that you generally hold the concept in your working memory for longer than typing. Which often times does not require the concept to be held in working memory at all because typing speed matches reading speed.

I also wonder if looking down at the page as you write vs looking up at the board. Constantly looking up and down is like a high frequency spaced repetition game, whereas the latter is a more motor/mechanical practice.

Excited to see this on HN front page! Unlike others, I didn't mind the CC + 30 day trial. 30 days actually feels generously long as I don't see many other projects offering that. I think it's ultimately right move for this project to live long.

After trying it for a bit on a very very large codebase(more than the context window supported, albeit I haven't done anything superr insightful yet), the code suggestion does seem better + faster than Copilot.

However, I'm not sure if the "completion" UX is the best way to enhance human programmers with AI. And within the completion realm, leaning on the speed, ie. inferencing on every keystroke, is not that attractive for myself. What attracted me is really the context length. So I'd provide much more examples of cross context code suggestion, similar to the 3js demo from gemini 1.5.

Back to the completion UX thing. I feel like often times seeing the completion pop up is a double edged sword. The moment it pops up, it distracts me from "outputting mode" into "evaluation mode" to see if the result is correct. If it's right, then great, you've saved me time. But for the times where it's wrong(which... was quite a bit for copilot), it's actually a net negative as I now have to re-enter the "outputting mode" and force myself to ignore the new output that will get generated as the new keystroke comes out. With Supermaven, this "switch" happens 10x more than copilot because of the speed as well.

Cursor/Zed with their CMD K code insert is obviously the other "big" ai coding UX.(along with chat) Personally I like them quite a bit and wish they had the speed and context window that supermaven is currently offering. But tbh, all of the UX's feel a little off at the moment...

Just my 2c' as an amateur programmer!

Love what you guys are doing. Is it possible to host custom models on fal?

When I don't wear contacts for a while(just normal glasses), wearing them throughout the day also makes me a bit "tired". Even though the "resolution" is basically the same. But after wearing them consecutively for a few days, it becomes like the same as glasses.

I wonder if there's something similar here going on. Since we rely so heavily on vision for everything(Especially balance). Any difference to our normal perception will cause "exhaustion" of sorts. But maybe wearing it continuously for days can cause our body to adapt to it?(Which, obviously is impossible for AVPs)

depends on what "in the end" means haha

I guess a good enough AI agent is essentially a human worker.

I wonder if all the work that's being put in right now by agent projects will become more or less "useless" similar to those specialized classification models before LLMs. Or will it be an AI with OK intelligence + 100 novel tricks/hacks that creates an Upwork level general agent.

Yea that's pretty cool too, I heard some restaurants are also doing a 100% voice LLM to take orders.

Transcription, specifically Whisper, is one of those ML models where the accuracy is basically on-par with humans. So, I really expect a lot more to come out of real time voice/LLM integrations.(the ChatGPT voice thing is a good glimpse, but very janky)

Good point, I should've defined this a bit more clearly in the post.

Honestly, I'm not toooo sure how to segment the term "agents", but in my mind there seems to be one realm for retrieval assistance. Ie. how do we make the ChatGPT-ish experience better. How can I better extract information I need from the collective human knowledge base. And another realm for letting the agent do things so I don't have to do it. Ie. "how can I get an Upwork assistant/Chief of staff/freelancer for cheaper and faster".

Nevertheless, editing the post now would simply create more confusion. Hopefully this discussion at least invites conversation about the conversation on agents itself haha.