Onavo

I don't think anybody hates AWS as a technology. It's mostly the rent-seeking (pricing almost-free bandwidth to an egregious level, nickel and diming users through fractional pricing calculations that are about as transparent as the US health insurance billing, predatory B2B vendor lock-in practice.) that gets on people's nerves.

When MBA (and YC startup) schools teach to build product with a "high switching cost", they do not have consumer's (in this case the developers) interests in mind.

The other issue is that outside of the core offerings (S3, EC2/Lambdas, the logging and queuing services), everything else seems to be in a perpetual state of beta with products shipped by interns. The situation is not as bad as Cloudflare but the bar's on the ground. Pre-LLM, services like Cognito caused developers no end of pain and suffering.

The two things that make or break an airplane are the engines and airframe. Everything else can be swapped out. As long as the airframe can support the engine, structurally and aerodynamically speaking, there are no issues. Wiring and avionics can all be easily replaced. The control systems for the plane can be reworked too.

GP is correct, this is a plane designed almost pre CFD, a clean sheet design would go much further than this. They are missing out on half a century of material science advances.

Maybe they can start on sandboxing next. Pi doesn't have a good sandboxing extension that offers Claude/Codex style auto approvals

Nah, the top quant companies and OnlyFans beat their pants off these companies for profit unit economics, especially per capita.

Do the Amazon numbers include AWS? I would certainly hope their ecommerce division is at least comparable to Walmart and Costco. AWS needs to be split out when analyzing Amazon because it distorts things.

A contractor CEO? I know it is extremely common for directors to be contractors for tax reasons.

Would you need a source map then for prod debugging?

I miss the YOLO (CNN computer vision model) days where one dude can publish a paper, completely disregard academic conventions, and yet push the field forward by leap and bounds.

Bet you they will use an external LLM world model to emulate the tools going forward. It's basically what's done in self driving research.

You are thinking from the point of view of an engineer, and risk falling into the rsync dropbox fallacy.

Start from thinking about product value add, and user behaviors first.

Imagine NVIDIA on steroids. Think bad documentation, opaque code, and everything being locked down. The customer gets squeezed for more money (especially egregious if you don't have enough scale they won't even bother engaging).

Sam at least can be understood. He's in it to make money and nothing else. The most dangerous men are people like Dario, who sit on top of their elitist towers and think they are right and will burn down the world in the name of righteousness.

Isn't that effectively the same as the blog post? You are just pushing the token filter to the sampling step.

The next gen (v15) is supposedly going to be around 10B.

> Could a cloud-delivered LLM figure out how to drive this route, based on those input data and given access to those output actuators? Looks like yes. Sure.

Well, if the massive cloud models that are generalized and have a world model that's good enough, you can just distill them into smaller models. As a point of reference, the current gen of Tesla FSD models only have 1B params. They are tiny by LLM/VLM standards.

Software and design costs are irrelevant in the age of LLMs.

In the olden days we call this classifier, usually assignment 2 of Machine Learning 101. BERT (well, GLiNER specifically) and diffusion are calling and want their Large Classifier Models back.

https://github.com/vllm-project/vllm/pull/57250

I think it happened many times to electrical lines due to lost campers.

Click bait title :/ I thought it's a machine learning article

Note you don't need a sandbox if you are not doing code execution. There are a lot of applications where inference only is sufficient e.g. web scraping websites that don't change frequently or OCR on scanned documents.

Code execution (usually TS/JS or Python) is useful most when you are dealing with truly open ended problems. It's the opposite of the use cases of most enterprise SaaS.

It's not limited to their libraries. All of their public facilities are like this. Even their public swimming pools (not allowed to use them if you are a visitor, their digital ticketing system requires the local equivalent of a SSN). The official excuse is that it's tax payer subsidized.

It's very much a classist country where you position in society is determined by your immigration status. Everything there is tiered based on whether you are a "citizen" (first class), permanent resident (second class), and you are in a permanent out group if you are not in the above two classes. And if you are a foreigner, whether you are a white "expat" or a Dubai -style indenture servant migrant worker. Valley tech bros are obsessed with that country but all that glitters is not gold.

Regardless there's value to know exactly what topics are off limits and for which platforms. Censorship needs to be measured and quantified (probably with vectors).

Is French different? Il est un médecin? Il est médecin? Ils sont médecins sounds fine.

To avoid vendor lock-in.

I think by one way they are referring to cruise missiles i.e. the engine is only used once or twice. Once during test fire, and once in production environment.

So...if you vibe code API shims Google can't sue your right? It would be clean room implementation by definition.

He didn't actually name his son Placeholder right?

I don't think anybody's denying that small models are doing distillation, the issue at hand is frontier lab vs frontier lab when it comes to big models.

Give me a proof they don't. Because from my observations they clearly do.

> even if I give them literal infinite time and all the subagents and internet access..

Don't use the word infinite in any CS claims. They can recreate or approximate monte Carlo tree search and it technically is still a correct solution in your framing of the problem so long they defeat you.

Anyone down to vibe code a competitor to ForeFlight?