WATCH: “Not even an hour!”
KANSAS Rep. @RogerMarshallMD leaves his own town hall after voters on both sides of the aisle express concern about veterans being fired from their federal worker jobs.
The AI models that “reason” are randomly guessing and checking against some deterministic external model.
Say you want to solve 3x + 9 = 4. A human looks at this, recognizes that numbers symbolically represent numbers and X symbolically represents an unknown quantity, then applies elegant deterministic rules to cheaply solve the problem.
The LLM doesn’t know anything. It has no system by which to define any symbols or rules other than how often certain symbols appear near each other. So the only thing the computer can do is hope the question is verbatim in the training data so there is a probability it will randomly select the correct answer. These are what people call hallucinations, which is misleading because they are natural and expected outputs from a stochastic model.
The desire is for the LLM to stop saying wrong answers. To do this, the labs are introducing reinforcement learning, where they take the deterministic model 3x+9=4 and feed the LLM answer into an error function like err = 3x+9-4. From there, the model then tries many values for x and returns the value with the smallest error. As the problems become more complex, you must randomly try more things, and compute balloons rapidly. This is why openAI had to spend massive sums of money to solve ARC-AGI, it was essentially brute forcing guesses against the training data until it got a match or it gave up.
In my world we call this dumb brute force, and it’s what we already use computers for. Throwing an LLM into the mix and trying to brute force every task on earth is one of the stupidest things you can do.
So ironically AI is getting exponentially dumber every day.
The real solution is to figure out how to create a network that can learn and encode symbols into a consistent world model, and do it efficiently. This will involve complete reworks of both software and hardware. And the fact that we are getting further and further away from this with current AI techniques is evidence that no one has any idea how to build it.
SNAP benefits for all poor Americans is about $150B a year.
AI benefits to a few rich assholes is now $100B a year. When it comes time to pay the piper and curtail government spending, don't forget who the true welfare queens are.
During my 60+ years of watching markets I’ve seen a lot of frenzied bursts of speculation in stocks with truly nebulous fundamentals and/or highly-dubious financials – but never in something that has no fundamentals at all. $TRUMP.
Scripture says: “I have been young and now I’m old yet I have never seen the righteous forsaken.”
After all these years serving you, the American people, I have not seen the righteous forsaken.
I love you all.
May you keep the faith.
And may God bless you all.
It’s a damn shame Joe Biden isn’t getting a second term.
Watch him actually answer questions like an adult with depth—then compare it to Trump’s word salads. The downgrade is almost comical.
When I took office, the pandemic was raging and the economy was reeling.
From Day One, I was determined to not only deliver economic relief but to invest in America.
We passed legislation to rebuild our infrastructure, build a clean energy economy, and bring manufacturing back to the United States after decades of offshoring.
After decades of trickle-down economics, we've written a new playbook that's growing the economy from the middle out and the bottom up.
We’ve helped create over 16 million new jobs, achieved historically low unemployment, and seen a record 20 million new business applications.
The economy created 227,000 jobs in November.
Unemployment has been the lowest on average of any administration in 50 years.
This has been a hard-fought recovery, but we are making progress for working families.