@Allinoltech2016 @CPATaxTeam Negative, the 70k is the maximum irs limit (not including catch up). Mega backdoor allows you to contribute Roth in excess of 23k deferral limit by contributing after-tax money and converting to Roth. Only works in solo-k plans or plans that don't have non-highly comp Ees
@stevehou@profplum99 If you're going to write 8 paragraphs, there's no need to use abbreviations like ofc and otoh. We're not 13 year olds texting each other
@petermessana@rdd147 Curious question, if you're following a car and they drive off the road into a ditch what is your reaction? Cause most humans would follow that car into the ditch
A thought experiment regarding this push for "general" models of intelligence ie a model that can do anything.
Suppose I have define "anything" as, let's say, any type of VLSI silicon integrated circuit, using only processes and structures already available today. Let's further suppose I can compute every possible chip that could be made as x possibilities.
Now let's say that I want to make a model that predicts the performance of those chips for customers of my model to query. Assume we can speed up the time for a customer to predict chip performance by 100x. Say from a week of compute to 2 hours.
I can't compute every case so let's say I compute just enough to linearly interpolate all cases with less than 10% error. Let's say this can be done using x/10 chips. So I cut my simulation space down by 10x!
Let's say all users of these simulation tools are willing to accept 10% error and will gladly pay 10x normal compute cost for the 100x increase in speed.
Now here is the most important part. There are so many chip possibilities that will never be made that no amount of demand would ever get anywhere close to even 1% of all possible cases over the life of my business selling this output. In reality it is vanishingly small. But let's presume for the sake of argument that my customers pay to simulate 1% of all possible chip combinations that I simulate. Remember, they are paying me 10x for this even though I am delivering them an answer that is 10% less accurate than industry state-of-the-art.
My costs are then the cost of compute for a single chip multiplied by x/10. I sell 1% of the solutions at 10x markup, I have only recouped 10% of my compute costs. I have lost 90% of my money. This is because I have simulated many, MANY things that are not valuable to anyone.
This is the current approach that all current AI companies are taking with gen AI. To anticipate all future queries from customers before they are asked and compute them all. This is why general models are destined to bleed money no matter how popular they get. Currently, the largest LLMs are still losing money and their accuracy is still well below 10%. Making them bigger to try and improve accuracy makes them even more expensive because they must essentially predict all future outcomes of the universe to provide information to every query with enough accuracy.
So now that I have shown that on basic fundamental grounds that general models will never be economical, let's play a trick.
Let's go one step further and try to get more clever about predicting what customers might want by sampling all of the existing chips that have ever been made. We argue that most of our customers are going to want things only incrementally different from the past and not radically different, so this sampling will capture most of the chips people will want to simulate. Say this reduces our simulation space by 10x. We now have a model that provides customers a tradeoff of 10% error for 100x speed improvement, at which point we can break even on our compute spend.
The lesson here is that as models become less general, ie specialized models for very specific tasks, the fundamental compute required to train the models becomes more economical. Small domain, highly specialized models are ORDERS OF MAGNITUDE more efficient than general models.
This is why the world is already built on small domain specific models. One natural outcome of this simple fact is that there are autonomous robots in every supply chain in the world. Most of our supply chains are already automated, but there isn't a single humanoid robot. Not a single one. Why? Because the cost to make it general grows faster than the usefulness of the output for most people. Pepsi doesn't need their bottle capper system to dance an irish jig or flip burgers.
This is why, fundamentally, general models like LLMs and humanoid robots are doomed to fail. Simply doomed. Because the costs outweigh the benefits, and the more you scale, the worse it gets.
Gen AI and Humanoid Robots are scams.