With Google Assistant (old assistant) I could say "Hey Google, play daft punk" and it would start playing Daft Punk on Spotify.

With Gemini (new assistant) it says "sorry I cannot play music, but here are links to services where you can find Daft Punk albums".

Fortunately at the moment you can still toggle between them. I guess not for long though.

Optimistic updates on the frontend are probably simpler too.

The short blade on top is perfect for breaking the tape to open the box without damaging the contents. Then the mouth can be used for quickly breaking down boxes or cutting shrink wrap. You are just cutting tape, so the blade never wears out.

I cringe every time I see someone using a Stanley knife in a supermarket.

https://www.safeknife.com/

Taking SpaceX as an example, they have increased prices across all their consumer products over the past six months. But they definitely aren't short on money with Alphabet and Anthropic combined paying them over $2 billion per month.

Microsoft/GitHub lost out here as they were just repacking other people's products.

I basically don't use Facebook any more because of this. Opening the app shows me the most sensationalist, fearmongering and outrage bait content they can find. It's worse than news channels (I haven't watched TV for 20 years). I have auto play turned off, and every few months the setting gets turned back on.

The first few posts I see scrolling through:

- US adds mandatory tips ahead of world cup.

- Woman dies after being hit by Audi in city center.

- There was a huge queue for women's toilets at a tech conference.

- 50% off mattresses.

- Some influencer I don't follow bought 20 rolls of 3M tape from Lidl.

Do I still have friends, do they still post stuff? I only see them in the reels/stories section at the top.

The only reason why I still have Facebook is because of groups. It's the main 'groups' tool people use in my country, so there are various local groups I am part of.

Instead of solving things yourself, you need to learn how to describe them in a way others can solve them. Otherwise you will just be fighting the instinct to just do it yourself.

https://marshallbrain.com/manna1

Comments URL: https://news.ycombinator.com/item?id=48321730

Points: 2

# Comments: 0

Malta has a population of 500k. Let's assume 100k people use MaltaGPT daily, and they send an average of 10 messages per day, so roughly 1M messages per day. That averages 694 per minute, but at peak could be 3-5x that, so let's say 3000 per minute. Usage will of course vary by day of week and time of day (they could partner with a Pacific island and share inference hardware).

Those 3000 messages per minute translate to 50 messages per second. Let's say average prompt input is 5k tokens, and output is 500. So 250k tokens per second for prompt processing (let's ignore caching for simplicity) and 25k tokens per second for output decode.

If we take a 500B dense model, that concerts to roughly 1 trillion flops per token. So we need 250 petaflops per second of prompt processing and 25 petaflops for output decode. So 275 PFLOPS in compute.

That may sound like a lot, however a NVIDIA DGX B200 machine (8xB200) has a compute of 144 PFLOPS at FP4. That is assuming 100% efficiency which isn't really possible, and we also need to factor in memory usage which we would be limited more by than compute. So let's say we'd need 10 of them. For an entire country to have a sovereign version of ChatGPT.

The cloud cost to rent one machine is around $50/hour, so that would mean our cluster comes to $4.8m per year. However the list price of a machine is around €400k, so the price to buy the cluster outright would be around €5m (you need the rest of the data center too), with operating costs of around €500k per year.

So per citizen: €10 upfront and €1 per year.

As I understand one of the reasons against using materials like stainless steel or other alloys for cars is that is is harder to work with, but most new cars today are written off - rather than being repaired - after even minor accidents, so that doesn't really seem like it's a realistic concern.

The backstory is that in the late 2050s when AI has its hands in everything, humans loose trust of it. There are a few high profile incidents - based on AI decisions -, which cause public opinion to change, and an initiative is brought in to ensure important systems run hardware and software that can be trusted and human reviewed.

A 16bit CPU architecture - with no pipelining, speculative execution etc is chosen, as it's powerful enough to run such systems, but also simple enough that a human can fully understand the hardware and software.

The goal is to make a near-future space exploration MMO. My Macbook Pro can simulate 3000 CPU cores simultaneously, and I have a lot of fun ideas for it. The irony is that I'm using LLMs to build it :D

We sell digital signage, and a host of related products, that create interactive retail experiences. This is powered by a small stick, that you plug into the back of any TV. That stick is Intel hardware, which runs Linux, various Docker containers, and the main UI for our service, through Electron.

The software that runs on it hasn't really been touched for a few years, and we are running into issues. The main issue is video decoding, where we cannot get hardware acceleration to work in Electron. For 1080p we've been able to ignore it, but we want to support 4K now, where things don't work without video acceleration.

We are looking for someone who is an expert with Electron + Linux + dealing with driver issues between them. If things go well, there will be more work down the line (and if you want, a full time position).

If you fit the bill, but are located elsewhere, please still get in touch. These locations are preferred, as it'll make shipping hardware for testing to you easier, but for the right candidate we can make an exception.

Email me to apply: luca@blissfulsystems.com