Onavo

That's a quick way to lose all of their customers. Inappropriate messages offering the illusion of privacy was the entire point of Snapchat.

A lot of of the library-as-a-service startups doing excel/PowerPoint/PDF editing libs are gonna die with how ubiquitous and powerful modern LLMs are. I expect a lot of such companies to write sob blog posts similar to what the tailwind founder did about having to lay off most of their team and reinvent themselves for the agentic era. It's a sad time to be a dev right now.

> If we stored the edges (links) and nodes (pages) separately, rather than requiring you to blindly run a node's code just to discover what its edges might be, then you could skip the prediction and instead pre-cache the next hop for all edges just in case you follow one. You could even do this to two or three hops.

Welcome to Next.js

> One thing I regret is the experiments as well. In physics and chemistry they were kind of a blur.

Well, a lot of places would penalize you if your data is incorrect, so most of the focus ends up on execution rather than the why. If you were angling for a high GPA for whatever reason, you cannot afford to experiment too much during lab work. It's another case entirely of course if you are conducting research.

Regardless, he chose to fold.

>I expected to read about legal threats against the folks collecting this data and am glad that those didn't materialize.

That's because the author folded over and played nice. Imagine if he exercised his first amendment rights.

The Chinese models also like to cut corners on stuff like science. Their scores on stuff like biotech and scientific knowledge is far from ChatGPT unfortunately. (Claude is pretty good but it just refuses all prompts).

> Cartels aren't known for massively increasing production to satisfy all demand.

Don't worry they will leave money on the table and lobby for protectionism when china comes and eats their lunch.

Simpson's?

So this is the Dutch equivalent of a cottage?

Not sure if you realize it but biology is a lot harder to regulate. Most of the secret sauce in biotech is in the manufacturing process, not the reagents itself.

I will give you a classic example, the Lipid Nanoparticle used for the delivery of mRNA vaccine is the real lynchpin of covid vaccines. The media never talks about it. Any biology student knows how lipid bilayers work (they are the basic building block of cells, it's like asking a programmer what a terminal is). But the process of manufacturing those lipid nano particles at scale was the main bottleneck for delivering mRNA to the target cell. That was the trade secret. It's not sexy or hot like mRNA so the media doesn't talk about it. It's why many developing countries tried to make their own vaccines and failed completely. The mRNA was the easy part but what use is a bunch of mRNA if you can't get them delivered to cells?

Any biology student (even high school ones) can describe how a cell membrane works. But designing a man-made one that can be created at scale and be guaranteed to fuse with mammalian cells is mostly a matter of trial and error. It's a bit like finding viable pseudo linear memory use alternatives to the basic attention architecture in machine learning. Every lab has their own take on it.

With AI nowadays the search space of viable biotech ex nihilo experimentation is much smaller given that you can just prompt it to give you plausible protocols.

A more interesting question is...does differential equations based models like mamba/state space models perform better on this sort of physics problem than pure transformer LLMs?

I hate to ask this..but how exactly do you plan to turn this into a scalable business beyond being yet another Oshkosh? Sure it's profitable but there's a reason hardware manufacturing is generally outsourced to the lowest bidder. Outside of targeting specific sectors e.g. defence, how else do you plan to remain competitive globally? Manufacturing hardware (especially if you are not Apple) is generally a capital intensive, low profit margin business that is extremely sensitive to labor costs. The fact that the past two decades of growth in the US is dominated by tech companies is precisely because software and the associated network effects are able to drive disproportionate amount of growth, unlike hardware. The further down the value chain in hardware the more commoditized you are.

It doesn't do auto approval.

The biggest issue with Pi is that they don't have proper sandboxing with auto approval. Most solutions are third party and half baked. You have to either choose between just auto approval (no sandboxing), or a Claude/Codex style sandbox but manual approval.

You can do live parallel routing and blue-green migrations transparently. In your docs you already describe how to do it but the user has to do it manually at the application level when you can do it just fine at the DB load balancer/router level.

Interesting, last I checked PlanetScale still doesn't have in place Postgres version updates.

Vercel does. They don't get enough credit for it.

Zeit is good. Glory to Zeit.

I wish mise has better secrets management or allows for split config e.g. a repo wide config that's to be shared and a personal config that can override whatever's necessary. It's difficult to commit mise.toml because the [env] section almost always contains sensitive variables.

The colo pricing is a fraction of a fraction of what AWS is charging. There's a reason that piracy video and file hosting sites can survive on just ads alone. If you try building something like YouTube on AWS you will go bankrupt in two days. The big cloud bandwidth pricing is the biggest scam ever. They don't want to charge up front for the human cost of running servers so they bake it an exorbitant overhead into the variable consumption cost hoping most web devs are too stupid too worry about it (it's usually Somebody Else's Problem).

I think every major email hosting provider (that's not a dumb IMAP only platform) has some variant of this.

Considering how ridiculously verbose their library is compared to React/Vue it's a good thing we have LLMs to deal with the boilerplate now.

I think the biggest issue is that cloud providers turned something that's a cheap commodity (bandwidth) into something that needs to be nickel and dimed. There's a reason AWS is Amazon's most profitable division and it's certainly not because they offer the best value towards customers.

It being vibe coded aside, does it support screenshots? I noticed the daemon mode has a lot of weird limitations too like not being able to return html.

Also you don't need anything fancy when washing them, their surface is just glass.

Can it handle EVs safely? The bottom battery often causes...issues if you don't have a lifting puck to spread the weight.

Also it would be nice to support browser cookies and flows too.

But your average Cessna lawnmower engine is perfectly startable without computer assistance.

The challenge is all legal. Every single AI shops is training on pirated data corpuses and grey area scraped data so obviously not a single company is willing to reveal their data.

Well, the snobby pure maths folks look down on applied engineering math ;)