So, by definition, driver of a tram is guilty for the accident the same way the driver would be if he had run into another vehicle. Tram hit the bus from behind.
“Tramways” are a specific type of road dedicated for trams, separated by dividers. “Tram lanes” are regular roads with tracks that are for exclusive use by trams, usually for part of the day and shared with cars for the rest.
The majority of the tram network is on roads shared with regular vehicle traffic (ie. neither a tramway nor a tram lane).
When trams are moving on shared roads they have no right of way special to them EXCEPT at roundabouts.
(I also know a PTV tram operator who lost their job for rear ending a car that suddenly stopped to turn right)
Anyhow, fair point - I principally jaywalked and /or rode a beat up mountain bike getting about Melbourne in my time there & took the road rules more as an aspirational guideline than as Commandments from the Victorian TAC.
Except trams have, as far as I know without exception, specific traffic rules around them. A tram simply does not and cannot stop as quickly as a car (or indeed a bus) can. The reason that that's largely accepted is because the tradeoff is also that a tram can't really make any sudden unexpected motions at all.
If you get hit by a tram it is because you are in a place where you should have known the tram would hit you. The things are on tracks.
local tram drivers union representative from Goteborg where it happened:
"The question is whether, for example, they’ve entered the correct right-of-way rules for trams? All vehicles are required to yield to trams, with a few exceptions, and that could be what’s “triggering” the system,"
Everything that the parent comment said I tend to agree, and I lived in Europe for a long time. One thing he left out, and which is also very important, is a high level of corruption. Your success is defined by a function who you know and not what you can actually do - EU funds being a prime example but even without them every day work is the same.
And they obviously don't hire the right people. Reasons can be many-fold. One possible explanation is that there's not many talent left on the market, and most have been already picked up by other AI labs paying more $$$ while offering more exciting work and more exciting trajectory at the same time. Another possible explanation is that there is enough talent on the market left but their recruitment process doesn't allow them to recognize those people, hence it is broken. When I look into their job postings, I tend to give higher chances for the latter.
You lay out some good arguments but I agree with both: the models relative to few years back really did become the commodity because today you could take the non-frontier model, maybe self-host it or pay the much less price per M tokens to get the performance of a ~2-year old frontier model. At the same time I do think that we are getting into the monopoly/duopoly/tripoly with the frontier models for all the reasons you already mentioned, and this scares me a little bit.
Lower intelligence LLMs can be a commodity, yes. But these won't make much money, if at all. At the end of the day, it costs the same to inference a 1T frontier model and a 1T free model.
OpenAI and Anthropic don't compete in the LLM commodity market. Hence, I had a problem with slide 22.
2-bit quantization? That's a lot of signal being removed. Considering how quickly the AI models are progressing in their capabilities (still exponential curve), I will not want to use the 2025 model in two years time. Similarly, how I don't want to use llama-3 or old Anthropic model from 2023 or 2024. Newer models are so much better that it makes it very difficult to ignore.
Once and if the advancements with the AI models slow down, only then IMHO it will become feasible to design the specialized HW for general-purpose consumption and general-purpose workloads.
Opus 4.6 was a 2025 model and many people (myself included) feel that if that's where models peaked, we won't be disappointed.
Even at 2-bit quantization, DS4 is probably on par with a 2024 frontier model. You can run that today on local hardware, and at a minimum, local models are going to keep pace over the next 12-24 months. Even if they don't close the gap with frontier models, they'll still play an important role in the overall pipeline for cost, speed and privacy reasons.
That's without even mentioning the additional capability that something like a Taalas chip churning out 17k tokens/sec could unlock.
The rest of the comments, as of now, are unreasonably unsubstantiated and a little bit hostile. AFAICS it's a large effort combining some pretty damn cool ideas from database internals world - this must have been written by someone experienced. It reminded me of RocksDB immediately, and wanted to see the example of integration, but then I saw redis, memcached, and "financial" database workload/protocol implementations. Good stuff, especially the benchmarks.
The code was obviously written using LLMs and I don't say this because the code looks like it's been LLM-generated but because of the fact that no sane person would have been able to write such complex piece of software in the pre-AI era. To me personally it shows how things dramatically shifted in software, and how domain expertise along with the AI became not 10x but 100x multiplier.
Not really a meaningful comparison because you haven't defined how much of a temperature difference versus how much of a water flow difference we're talking about here. But for most people, at least if they take the most basic level of care to not use water at like 85 degrees or 110 degrees, then no, that's just not really true.
There's a lot of folklore out there that's lingered from the early 2000s espresso community where it was widely believed that temperature was the holy grail control parameter, but now with modern instrumentation and temperature probes, it's been pretty much debunked. Temperature stability throughout a shot makes almost zero perceivable difference in taste.
It takes brew temperature differences swings of around 5 degrees Celsius before people can start to notice any difference better than random chance, and almost a 10 degree brew temperature difference before it gets to the territory of 'ruining a shot'.
Meanwhile, very small differences in puck preparation, including micron-differences in grind size, or sub-gram-level differences in coffee quantity have profound differences in flow rate, which has a very strong affect on coffee extraction levels, which has immediately recognizable differences in the produced flavours that a trained palette can reliably detect. This is before we even start talking about channeling which has an enormous affect on the coffee.
Manual control of the applied pressure can and does allow skilled people to compensate for those differences in flow rate, and combined with very basic attention to brew temperature, does help shot consistency.
So much bs that I don't even want to go further into discussion, sorry. I say this as someone who has made several thousands of espressos on E61 group machine. I'll let you have your own opinion but anyone who has made more than a few espressos will immediately understand if and when the temperature drifted away. Pressure? I've made espressos at 6 bars and 9 bars. Makes literally almost no impact or whatsoever. You're right though that 5 degrees Celsius is probably about the right minimum amount when the espresso starts to change in taste, and there's remarkably many machines which cannot sustain the temperature in shot after shot workloads.
A good machine will always need time to heat up since the temperature stability, which is very important for getting good espresso shots, correlates with the weight of the device. For most machines this means 30-45 min or so no matter what the manufacturer is saying and in practice this isn't much of a problem once you plug the machine through a smart plug which you can program to turn on the device before you're getting up in the morning.
Secondly, adjusting pressure is almost a completely unnecessary feature so I'm not sure why do you chose to point that out as a major differentiator. 9 bars is just fine. In similar category goes the PID for adjusting the temperature. While on the paper it sounds cool in reality you will not use them 99% of the time. There's many prosumer machines which don't allow you neither of those and are still perfectly fine machines.
Do you know what is it that it allows for such a quick warm up? Small boiler? Saturated group? Maybe my comment is more relatable to HX machines then although I don't quite get how is it possible to warm up so quickly - the machine is still a 20kg piece and you can't beat physics with such large thermal mass.
250ml brew boiler with a small saturated group, and a separate 1.6L steam boiler. It's definitely aimed at a home environment but will easily keep up with making several drinks in a row. Compared to previous HX machines I've owned, it heats up much quicker and is more stable. I think not having a massive lump of brass like an E61 style machine helps a lot.
They have separate tanks for keeping water heated up to temperature wit a much smaller volume and don’t use saturated groups so the total hot mass is much smaller.
reply