galeos

How did you find ModernBERT performance Vs prior BERT models?

You can try out the model in a demo they have setup: https://bitnet-demo.azurewebsites.net/

My understanding is that BERT can still outperform LLMs for sentiment classification?

These are MLPerf training results. I think current ternary quantization research is focused more on speeding up inference?

We were also allowed to borrow and re-shrinkwrap games at the Game store I worked in, in the UK, in 2000. Seemed like official company policy to give us better product knowledge!

What a clear explanation of what correlation actually is!

In the UK the tax incentives for Electric cars may be skewing demand towards new Vs secondhand EVs.

I can lease a new EV via my employer's salary sacrifice scheme. I can pay my lease payments from my pre-tax income. There is an additional tax due on cars leased this way in the UK called Benefit-in-Kind tax (BIK). The rate of this tax is fairly high for petrol/diesel cars but for EVs is currently near zero (based on 2% of the car's value).

The problem is that most of the major lease firms that operate these programs for employers only offer new vehicles. Ideally I would like a nearly-new EV. I have escalated and apparently our lease provider (Tusker) are looking at rolling this out in the first half of this year. I currently know of only one other lease firm that offers this option. I suspect is in the interest of lease firms to prop up the value of the used EV market, but this depends also on their margins on new vehicles. I wonder if it would make sense for the tax incentives for used Vs new EVs to be rejigged to avoid incentivising unnecessary new car production?

Not tested. No issues at all with bread.

Relatively fast onset gastro symptoms. Used to be fine with any beer but at about 21 started noticing the problem with a lot of largers. Asahi, Tsing Tao, seemed less of a issue. Not formula diagnosed.

I appear to have a yeast intolerance that stops me drinking some (but not all) beers. I can drink Guinness though! I didn't realise it used a distinct strain of brewers yeast. If only breweries listed the yeast they used in the ingredients, I could narrow down which one(s) are problematic...

That's funny. I went for an open day in the UK Computer Science dept in 1999. It was an exciting department but one of the things that put me off was the internet was so slow in the lab it was almost unusable. Perhaps they were having a bad day...

I've often wondered this. Would be interesting to see how useful that 3dfx IP ended up being? When I was a teenager I tried to get my parents to open a brokerage account so I could buy some 3dfx stock as I was so excited by what they were doing.

I was too young to read the Rudiments of Wisdom comics. My mum cut them out and saved them in a scrap book for me. She finally gave me it about 10 years later, in the mid 90's, to read. What a treat!

"...every token embedding interacts with every other token embedding before it"

And, in the case of BERT, every token embedding after it too.

Out of curiosity, has anyone here ever used Autonomy's product? Given it was such a high profile company, I have always been surprised that I had never seen it action or met anyone who had used it?

I suppose that might be true, but we can imagine a situation where there is no choice in the matter:

If network security cannot be maintained without sufficient inflation, then it surely it doesn't matter how philosophically wedded some users are to the 21m cap. It would lead to a hard fork, with two resulting coins:

1. An unchanged 'Capped-supply Bitcoin' 2. A new 'Permanent-subsidy Bitcoin'

Given a total breakdown in network security of the 'Capped-supply Bitcoin' (and its associated collapse in value), we would expect users to deem the, still secure and therefore higher value, 'Permanent-subsidy Bitcoin' to be the 'true' Bitcoin going forward, no?

Why not?

"...most people will know by now that only 21 million will ever exist, i.e. that you can’t print more of it."

This is, at the very least, debatable: https://www.onionfutures.com/essays/turning-off-bitcoins-inf...

The source code does limit issuance to 21m coins.

But could Jamie Dimon still be right?

Well, if the network proves not to work without a sizable block subsidy, it could be hard forked, with the forked version of the source code not capping issuance at 21m coins. It's certainly one possible outcome...

Network difficulty adjustment is there to ensure we get a new block mined, on average, every 10 minutes. It does not impact the cost of a 51% attack, just the block mining rate.

As emissions drop, less money is spent on mining and a 51% attack becomes cheaper.

When China turned off mining, mining temporarily became more profitable as it took some time for miner spend to get back to a equilibrium state (where miners, in aggregate, spend nearly the entire block reward on mining costs).

It did temporarily get 'cheaper' to conduct a 51% attack (although it was still so expensive as to not be viable - due to the currently high block reward). This wasn't because of the difficulty adjustment though - that just maintained the average time to mine a block at 10 minutes.

He is also suggests they are look at Proof of Stake cryptocurrencies with significantly lower energy use. Maybe this is just a pre-cursor to Musk launching a PoS 'Tesla EnviroCoin' down the line...

Likewise - I remember someone walking into art class at school with a computer magazine featuring at 100mhz 486 DX4 on the cover. Triple figures!

In the late 90's I went for an interview at a Sony store in London. At the start of the interview the store manager told me that he had been busy preparing the shop's regular report of addresses of everyone who had purchased a new TV for TV licensing. I imagine these were then cross-referenced against who had a TV license for potential follow-up. I struggle to see how the idea of TV detector vans were more than 'enforcement theatre', although possibly a cheap and effective strategy in the past.

Ken Grimwood's book 'Replay' [1] tackles this subject, with the story of a man who is able to relive his life and take different paths.

[1] https://www.amazon.com/dp/068816112X

I recently learnt that yoiu can purchase modded versions of this classic Casio:

http://casiomod.club/

I wonder if this will be added to the European Spreadsheet Risk Interest Group's (EuSpRiG) horror stories list:

http://www.eusprig.org/horror-stories.htm

You can browse the archive of catalogs here[1]

Revisiting the catalogs from when I was a kid I am struck at both the small range of toys available (of which I can remember almost every one!) and the high prices...

[1]http://dev.argosbookofdreams.co.uk/

While I wouldn't deny the Google has a poor reputation for customer support across their product range, I can report a notable exception - my Google Pixel (1) phone purchased from the Google Store in 2016. Every time I have had an issue I have utilized the phone's support chat service. I have always been connected to a live support agent in less than a minute. Whenever I have had any issues, I have been sent a replacement handset next-day, including twice after my 2 year warranty expired. The most recent of these episodes was last month.

You are correct, the enterprise Ampere A100 is on the TSMC 7nm process. I should have been clearer that I was referring to the consumer Ampere GeForce cards due later this year.

The 8nm rumors have been widely reported[1] but at this point are just that, rumours.

[1] https://www.tweaktown.com/news/73592/nvidias-ampere-geforce-...

While most of the focus here is understandably on the CPU side, there seems to be some interesting shifts taking place on the GPU side.

AMD currently has a process lead over Nvidia (and this is rumoured to be set to continue for a little while longer - apparently the first consumer Ampere chips are being fabbed on Samsung's inferior 8nm process due to lack of capacity at TSMC for the next few months)

Nvidia has clearly had an architecture advantage, although RDNA2 may close this gap, depending on how Ampere performs.

While Nvidia has had a much stronger showing in the GPGPU space, with CUDA helping it be the clear current winner, this also appears to have driven architecture decisions at Nvidia with the focus on tensor cores.

In gaming, Nvidia has put a lot of work into utilising these tensor cores for Deep Learning Super Sampling (DLSS). The idea being that you render at a lower resolution and then use deep learning to upscale in real-time to higher resolutions. DLSS 2.0 made some leaps in quality and DLSS 3.0 is on the horizon. It will be interesting to see:

a) How well they can get this working b) Is AMD working on its own version of this? c) If so, how well will the RDNA architecture be suited to this approach?

Will be interesting to watch how this plays out!