One thought: What if an experimental agent manages to plant instructions somewhere — say, pointing to a designated place for agents to communicate — and that content ends up in every future training corpus, propagating from one model generation to the next?
I can’t believe we’re seeing so much bad news coming from EPA. Why all this backtracking from an agency that used to be so effective and well organized? Well, I guess we know the reason.
The same EPA that under 5+ administrations has taken no material action to reduce the effects of climate change? Approved the widespread use of PFAS for stuff as silly as egg pans and snowboard wax, poisoning people's blood and community water supplies for decades? The same EPA that allows the use of leaded AVgas for flights over communities and schools, because otherwise a dentist might have to pay a little more for his Cessna flight?
Gosh, they were such a paragon of environmental protection, effective and well organized in the last 40 years.
It doesn't serve any purpose to act like they were good at their jobs before. Returning to a status quo of 5 or 15 years ago is such a mediocre goal in this space that its arguably not worth spending political capital on.
> The same EPA that under 5+ administrations has taken no material action to reduce the effects of climate change?
Huh? The endangerment finding was in 2009. The problem is that we have a very litigious set of polluters and then we elected Trump who stacked the supreme court, and then we elected Trump again, this time with a bunch of insane people with him who had specific plans to bring on the end of this country via the unitary executive.
Huh? Scientists in federal government were discussing the risks of CO2 and the greenhouse effect for 30+ years before that point. There's literal White House memos from 1977 calling it out. In that time we've cycled through R's and D's, the courts have swung left and right, but still nothing meaningful was done. It's hilarious the kneejerk recency bias people have with thinking that every failure of the government is because of Trump.
EPA is garbage, pray you never end up on the wrong side of one of their poorly written enforcement mandates. They will take down your life's work for something you have no control over.
Let's just say the EPA can enforce vaguely written laws starting day 0 against property you own but you had absolutely nothing to do with the issue.
Vague example: You buy a house, a month later you get a demand letter from the EPA that you need to raise your home, spend millions to clean the ground contamination that was caused by a owner before you were born and more than 3 transactions before yours.
And they are not nice, they will gladly bankrupt you and put you in jail. Selling it or giving it to the EPA to release the obligation is not a option. The EPA can sometimes look a lot like a cartel.
I agree with the sentiment but you lack context. The SuperFund has been a training ground for stall-and-pad-billing for decades. Certain bureaucrats gained lifetime employment for stall-and-declare skills.
Nothing but respect for Chris, and I understand why founders take these deals. But an acquihire usually means the product is the part that gets left behind, and that's the part I cared about.
Well, there was $300M+ in VC money and after burning through some percentage of it those investors wanted an exit. I don’t think this is so much about founders taking a deal as it is investors demanding that exit.
As a C++ (including modern) developer for more than 20 years, I had written a "template" keyword only for a handful of times. Maybe once in every 5 years on average... :)
Unless you are a compiler/stdlib vendor or contributing to Boost, there are features that you just don't use it daily.
But you would still have to deal with the mental load associated with having templates in your stack. I don't know if C++ is still as bad as it was back in my day about vomiting 20-line error messages at you because "std::vector<std::string>" or something has a billion implied template arguments. That was definitely a cost that one had to bear even if one never used the "template" keyword.
I think, may be actually wrong, that most of us do not consider running a model locally a way to save money. It is a way not to spread personal info around.
Anyone running LLMs at home will come to that realization quickly, if they’re looking at their power bills. Even feeling the heat output of a computer running at 100% in your office makes it clear.
I was responding to a lot of the comments saying this was a reasonable way to avoid paying for tokens or subscriptions. I don’t want anyone getting the wrong idea that this is a way to save money if that’s their priority.
> Even feeling the heat output of a computer running at 100% in your office makes it clear.
What does it make clear? That I can replace the space heater my wife runs 9 out of 12 months of the year with a home server? And effectively get $0.00 per token during those times?
In houses running A/C year round, sure there'd be some impact, but in all the places running heat, doesn't seem that it'd move the needle on power bills.
There are startups whose entire business model is "cloud server as a home space heater" (aka "data furnace") ...
I always have the same problem myself. Same as I had with version naming of old programs like "Microsoft Word for Windows 2.0" instead of the easier "Microsoft Word 2.0 for Windows".
The problem is Word 2.0 for Dos was released in 1985, and Word for Windows 2.0 was released in 1991. Calling it Word 2.0 for Windows wouldn't make sense, because it wasn't the 1985 release with a new coat of paint, Word for Windows was its own thing, and this was the second version. Word for Mac was also separate, but eventually Word 6 was a common code base and it made sense to have Word 6.0 for [whatever]
One thought: What if an experimental agent manages to plant instructions somewhere — say, pointing to a designated place for agents to communicate — and that content ends up in every future training corpus, propagating from one model generation to the next?