londons_explore
londons_explore on Hacker News
On the face of it, running a model to a specific state, looking where is makes a mistake, and then figuring out which weights need to be jacked up to get the desired outcome doesn't seem like an impossible task...
And I've seen a worrying number of those fail after decades too. In cases I've seen, the fact that load is uneven between the strands usually seems to play a part - some always start to snap or pull out before others.
I had good luck with telling claude to rewrite my laptop bios to enable various power modes which were theoretically possible but disabled by firmware, like powering a fullsize screen from the laptop battery over usbc.
Maybe I'll tell claude to rewrite the car firmware too...
Mr boss man needs to uncork Google's 'labs' culture.
A good solution to this is to have a "break glass" system. Ie. access permissions get granted to the developer directly in an emergency by going through a set of steps available beforehand. All requests by that developer are then allowed, limited to 10 kilobytes of results per minute and logged.
For example, imagine you run a bank, and a rare code bug means that at exactly midnight on new years eve all bank balances have $1B added to them.
Your oncall takes down the production servers to to reduce losses and uses the break glass system, by getting approval from n other developers, allowing them direct database access. They use that access to count(accounts) with a balance around $1B or more, see it is everyone, and subtract $1B from them all, before bringing production back up with an older software release. Downtime = 2 hours.
In the morning, all the developers actions are reviewed.
If the developer turned evil and trying to dump the database, the n approvers and 10 kilobytes per minute would reduce data loss. But the break glass system still lets the developer make fairly major emergency fixes to reduce losses or downtime.
It all depends how many billions are going to paying salaries...
To be honest, if I were an LLM provider I would be wiring up my release process to be fully automated so within 10 mins of the last bit of fine-tuning it got released and a blog post posted.
The equivalent of a GPU but for bio lab experiments. It should be able to do all the basic lab operations of mixing chemicals, heating, cooling, various tests etc tens of thousands of times per second.
Then you hook many AI's up to the automated lab and watch science happen far faster.
That isn't the case in Shenzhen.
16 bit windows wasn't complicated, and a simple VM and API translation layer could have kept compatibility going back 30+ years for not many engineering hours.
Yet instead they dropped 16 bit windows support and probably spent more engineering hours to support all the ancient software trying to update, and users who held back upgrading windows because they needed said ancient software, etc
I believe the main difficulty is making large and pure diamonds. Making small low purity diamonds is cheap and easy to scale.
If you hold the key to saving a million people's lives, yet you decide to let those million people die so you can take the stage at a prestigious conference, are you a good person?
Curing disease is literally a race, with people dying every minute of every day. In any other situation, labs would be running night and weekend shifts and not a second would be lost.
Also, I live in a place where that was probably many months pay to the auditor.
Made my tax audit rather painful.... "I posted a joke on the internet about binary numbers not being worth anything and someone gave me $10k of bitcoin that I'm now declaring - is it income for the work of writing the joke, or a gift? The sender likely didn't know who I was, and was likely overseas but I cannot prove who they were or which country they were in".
At least some of this will be AI companies buying and scanning every book available...
Plenty of companies can donate 0.1 engineers to work on php, if for no other reason than to make php better for their own use cases.
And I assume there are also a high bandwidth comms channels, and this is just the backup which is presumably tested each day.
Presumably they don't want you leaking your conversion transcript to a competitor.