jimmyl02

The argument the article makes is it's not that different. Jev's argument is that they have trained the model to better output probabilities (which is not necessarily a training object of LLMs but we don't actually know that)

Ultimately Jev claims to have a data advantage which is likely where the future lies. They'll have a unique edge in improving general purpose classification / decisioning.

The architecture changes and systems improvements being brought into LLMs is so awesome to see. It really feels like this is now a systems problem where a defined goal is set then systems optimizations are made around the model architecture to solve it.

Underlying it all is that any architecture can be trained to the same convergence just difference in compute utilization both in training and inference

What's interesting is that we still read AI's output and that's still the primary interface between a human and the AI. Implicitly that means that writing is important and I think we're already starting to see this with the complaints around how Claude "talks"

It's also interesting how Clickhouse / Starrocks can now also act as a query planner and executor on top of non-native formats (ex. Iceberg).

I assume the native formats will always be faster / more optimized but the need for Trino as a separate executor while running either of these databases seems to be close to gone.

HTAP is here! It seems like these hybrid databases are slowly gaining adoption which is really cool to see.

The most interesting part of this is the improvements to transaction handling that it seems they've made in https://github.com/alibaba/AliSQL/blob/master/wiki/duckdb/du... (its also a good high level breakdown of MySQL internals too). Ensuring that the sync between the primary tables and the analytical ones are fast and most importantly, transactional, is awesome to see.

It's very powerful to have ownership over a framework that many developers are familiar and like!

It might not be clear just yet what the path to monetization looks like but an easy example would be deeper integrations with the rest of the Cloudflare ecosystem (for example allowing R2 to be easily along with something like duckdb to live in a world of truly local analytics or something like that).

It seems like these great open source frameworks need to monetize by building a platform around the product but these days it's hyper competitive (ex. Vercel, Cloudflare) and it's hard to get started without an incredible differentiator. So, while monetizing independently as a company might be difficult, Astro can provide a lot of value to the rest of the Cloudflare ecosystem.

Astro is my favorite framework for static sites. It's hard to describe but it just "makes sense". You don't need complex build setups and you can get the best of all worlds with great SSG capabilities + can easily express things with React when needed for example.

I wish there was another path to monetization besides joining a larger company but I'm happy that the team will get to continue building out an amazing framework!

Clickhouse's full announcement is here https://clickhouse.com/blog/clickhouse-raises-400-million-se... and I think another big piece is directly integrating postgres into their ecosystem.

It seems like an expansion play from their team and their end vision as both a platform (clickhouse + postgres) and product (observability) seems to be pretty good combo that fits hand in hand.

Usually there are nuggets of wisdom in lists shared like this but I feel like every lesson shared here has immense value.

> "remain skeptical of your own certainty" > "Model curiosity, and you get a team that actually learns."

These are two lessons that typically require battle scars to learn. For such big ideas to be summed into two sentences is pretty remarkable and puts to words lessons I wish I knew how to share. Amazing article, thanks for sharing!

It seems like this might be one of the biggest vulnerabilities in recent times...

The default react / nextjs configurations being vulnerable to RCE is pretty insane. I think platform level protections from Vercel / Cloudflare are very much showing their utility now!

if it increases topline metrics like watch time it's probably hard for them to justify removing it. a change this big seems like it was probably a/b tested and did move metrics significantly?

pretty interesting how datadog's uptime tracker (https://updog.ai/) says all the sites are fully available.

if that's true then it's a sign that Azure's control / data plane separation is doing it's job! at least for now

I think reve (https://reve.com) should be in the running and would be very curious to see the results!

always wondered at what scale gossip / SWIM breaks down and you need a hierarchy / partitioning. fly's use of corrosion seems to imply it's good enough for a single region which is pretty surprising because iirc Uber's ringpop was said to face problems at around 3K nodes.

it would be super cool to learn more about how the world's largest gossip systems work :)

in the super public consumer space, search engines / answer engines (like chatgpt) are the big ones.

on the other hand it's also led to improvements in many places hidden behind the scenes. for example, vision transformers are much more powerful and scalable than many of the other computer vision models which has probably led to new capabilities.

in general, transformers aren't just "generate text" but it's a new foundational model architecture which enables a leap step in many things which require modeling!

I think the conclusion isn't as simple as "foundation model companies will just build the features of all downstream products" because focus and priorities play a big part.

If that were the case a simple example is much of software services we see today (and provide real tangible value) wouldn't exist as it's theoretically just updating data in a database.

Cannot +1 this enough! Joining a team you respect and seeing how they operate gives you a really good baseline to work off of and take what you like and modify what you disagreed with.

You'd be surprised how many times you can "iterate and fail quickly" only to end up at an established practice some other shop has been doing for years. It is important however to understand the why behind the decisions as otherwise you're no better than just figuring it out yourself

Honestly this is a super respectable culture. I would think many leaders want to build teams with this mindset but find it incredibly difficult for one reason or another

This is super fun! The apple maps look for the "Find My" feel was a really nice touch

This seems like a similar and more feature complete / polished version of JSON RPC?

The part that's most exciting to me is actually the bidirectional calling. Having set this up before via JSON RPC / custom protocol the experience was super "messy" and I'm looking forward to a framework making it all better.

Can't wait to try it out!

This metric is typically tracked internally and probably wouldn't be as public because it could indicate how "buggy" a product is. An easy way to measure this is time spent taking incidents from open -> mitigated -> resolved and treating that as time spent * engineers for amount of impact.

The tricky part would be measuring time spent on hardening and making the business decision on how to quantify product features vs reliability (which I know is a false tradeoff because of time spent fixing bugs but still applies at a hand wavy level)

this being the 2nd large compromise of the week is not boding well from the NPM ecosystem...

supply chain is and has been the new gold mine for bad actors it seems

This is awesome! It seems to line up with the idea of agentic exploration versus RAG which I think Anthropic leans on the agentic exploration side of.

It will be very interesting to see which approach is deemed to "win out" in the future

I understand the concerns around a central gatekeeper but I'm confused as to why this specifically is viewed negatively. Don't website owners have to choose to enable cloudflare and to opt-in to this gate that the site owners control?

If this was cloudflare going into some centralized routing of the internet and saying everything must do X then that would be a lot more alarming but at the end of the day the internet is decentralized and site owners are the ones who are using this capability.

Additionally I don't think that I as an individual website owner would actually want / be capable of knowing which agents are good and bad and cloudflare doing this would be helpful to me as a site owner as long as they act in good faith. And the moment they stop acting in good faith I would be able to disable them. This is definitely a problem right now as unrestricted access to the bots means bad bots are taking up many cycles raising costs and taking away resources from real users

unfortunately sandboxes also have bugs which allows for RCE. this is typically called a "full chain" because it chains together a series of bugs (initial decoder exploit + sandbox escape exploit) to gain full RCE

I thought the reaction was more so that you can train SOTA models without an extremely large quantity of hyper-expensive GPU clusters?

But I would say that the reaction was probably vastly overblown as what Deepseek really showed was there are much more efficient ways of doing things (which can also be applied with even larger clusters).

If this checkpoint is trained using non-Nvidia GPUs that would definitely be a much bigger situation but it doesn't seem like there has been any associated announcements.

This is great perspective about how assumptions play out over longer period of time. I think that this risk is much greater for free third party services for critical infrastructure.

Someone has to foot the bill somewhere and if there isn't a source of income then the project is bound to be unsupported eventually.

this is awesome but I'm wondering does pgdog plan to handle high availability scenarios (multiple frontend proxies)? I know this can lead to much more difficult problems with consensus and handling split brain scenarios.

if not, what is the approach to enable restarts without downtime? (let's say one node crashes)?

I think there is no way to tell and we can only see with more research and time. One nuanced part that might not be clear is the transformer was a huge part of what made traditional LLMs scale.

With the diffusion transformer and newer architectures, it might be possible that transformers can now be applied to diffusion. Diffusion also has the benefit of being able to "think" with the amount of diffusion steps instead of having to output tokens and then reasoning about them.

I think it's hard to tell exactly where we are headed but it's an interesting research direction especially now that it's somewhat more validated by Google.

cool initiative! building apps locally without the cloud services has definitely been quite challenging both to setup and maintain when I've done it before