subj
The council did not approve
But seriously, the best relationships are those where your conflicts do not spread to all your friends and acquaintances.
subj
Crazy. Maybe will help keep top quarterback recruits coming!
Until yesterday, no company worth over $500B had ever gained more than 25% in a single trading day. Then came Oracle. In a move that defied both gravity and historical precedent, Oracle stock surged 40% today, adding over $300B in market cap overnight. The company now hovers just shy of the trillion-dollar mark, and Larry Ellison - armed with a 41% stake - woke up as the world’s richest man, suddenly $100 billion wealthier.
Yes, Oracle. The perennial punchline of “legacy software.” The company most of us had filed away in the footnotes of tech history is suddenly the market’s cool kid.
For those paying attention, this moment has been years in the making. Oracle’s pivot into cloud and AI wasn’t impulsive - it was deliberate, capital-intensive, and decidedly unsexy. They didn’t chase developer mindshare; they banked contracts. And those contracts just hit the ledger all at once.
➰ The Q1 revenue headline - $14.9B, up 12% YoY - wasn’t what lit the fuse.
➰ Even IaaS revenue at $3.3B, up 55% is strong, but not frenzy-worthy.
➰ The magic number was buried deeper: $455B in Remaining Performance Obligations (RPO), up 359% YoY. That’s nearly 8 times Oracle’s current revenue run-rate, a backlog so large it borders on the surreal.
RPO isn’t a flashy number. It doesn’t trend on CNBC tickers. But in enterprise software, it’s gospel. It represents revenue already won but not yet recognized. In plain English: Oracle just told Wall Street, “We’ve already signed nearly half a trillion dollars’ worth of business. All that’s left is execution.”
Oracle expects cloud infrastructure revenue which came in at $3.3B this quarter to hit $18B this fiscal year and ramp to $144B within four years. They noted that “most of the revenue in this forecast is already booked in our reported RPO”. It’s less of a forecast and more of a countdown at this point.
The market isn’t just reacting to a quarter. It’s reacting to a company that rewired its DNA and is now producing receipts. In a space dominated by Amazon Web Services (AWS), Microsoft Azure, and Google Cloud, Oracle carved out an edge not through branding or developer love, but through being the only one willing to say yes to what AI-native enterprises actually wanted: custom infrastructure, multi-cloud deployments, sovereign regions, long-term capacity, and massive scale contracts.
What we witnessed today is the rarest thing in markets: a narrative inversion. Oracle went from legacy to legend not by shouting louder but by building slower, selling longer, and letting the numbers speak. The company that once stood for on-prem databases is now one of the most valuable cloud businesses in the world.
TikTok and Twitter are obsessing over the ‘Great Lock-In’ without agreeing on what it means. Oracle just showed the only version that matters: half a trillion in contracts, signed and sealed. King of the Lock-In.
Yes, Oracle. The perennial punchline of “legacy software.” The company most of us had filed away in the footnotes of tech history is suddenly the market’s cool kid.
For those paying attention, this moment has been years in the making. Oracle’s pivot into cloud and AI wasn’t impulsive - it was deliberate, capital-intensive, and decidedly unsexy. They didn’t chase developer mindshare; they banked contracts. And those contracts just hit the ledger all at once.
➰ The Q1 revenue headline - $14.9B, up 12% YoY - wasn’t what lit the fuse.
➰ Even IaaS revenue at $3.3B, up 55% is strong, but not frenzy-worthy.
➰ The magic number was buried deeper: $455B in Remaining Performance Obligations (RPO), up 359% YoY. That’s nearly 8 times Oracle’s current revenue run-rate, a backlog so large it borders on the surreal.
RPO isn’t a flashy number. It doesn’t trend on CNBC tickers. But in enterprise software, it’s gospel. It represents revenue already won but not yet recognized. In plain English: Oracle just told Wall Street, “We’ve already signed nearly half a trillion dollars’ worth of business. All that’s left is execution.”
Oracle expects cloud infrastructure revenue which came in at $3.3B this quarter to hit $18B this fiscal year and ramp to $144B within four years. They noted that “most of the revenue in this forecast is already booked in our reported RPO”. It’s less of a forecast and more of a countdown at this point.
The market isn’t just reacting to a quarter. It’s reacting to a company that rewired its DNA and is now producing receipts. In a space dominated by Amazon Web Services (AWS), Microsoft Azure, and Google Cloud, Oracle carved out an edge not through branding or developer love, but through being the only one willing to say yes to what AI-native enterprises actually wanted: custom infrastructure, multi-cloud deployments, sovereign regions, long-term capacity, and massive scale contracts.
What we witnessed today is the rarest thing in markets: a narrative inversion. Oracle went from legacy to legend not by shouting louder but by building slower, selling longer, and letting the numbers speak. The company that once stood for on-prem databases is now one of the most valuable cloud businesses in the world.
TikTok and Twitter are obsessing over the ‘Great Lock-In’ without agreeing on what it means. Oracle just showed the only version that matters: half a trillion in contracts, signed and sealed. King of the Lock-In.
subj
LLM and diminishing returns
We are doing the same thing with LLMs. Increasing the model size by 10x will only give diminishing returns.
You have a 3D bust and you want to approximate it as close as possible with a fixed amount of triangles.
Having 600 instead of 60 triangles makes a huge difference, but having 6m instead of 60k triangles doesn’t make a real difference other than being 100x more expensive to compute.
This is exactly the same of what we are seeing with LLMs nowadays. The bust is the data we are scraping from the web, and the triangles are the model weights. And we have the same problem: Increasing the model size by 100x will be 100x times more expensive but we will get improvements that are really hard to notice.
Why? The problem is inherently to the LLM structure.
Models spend the same amount of compute on filler words like “however”, “therefore” or variables when coding or doing math. No one cares which filler word to use, but if you even just flip a sign in a math formula it’s just wrong.
It’s funny how in the model has high entropy on filler words not because it doesn’t know, but because it’s not important which filler word to use (this point extends also to synonyms).
And if you don’t do good data curation, which is currently being neglected by academics and industry (the industry is slowly going more and more) than you will spend a ton of compute during training on things like filler words.
Where do we go from here?
First big step is data curation. Our current datasets are so massive that we rarely do more than a few epochs when training. So we get rid of the easy examples to shrink the dataset to give more compute to the hard samples. This might require something like curriculum learning during pretraining.
Many people thing we have LLMs because of the transformer architecture or some other model idea, but actually it has been data all along. First ImageNet, then web scraping, large scale human preference data, and finally data from verifier systems. You can have a LLM without the transformer architecture, but not without the data. (read https://lnkd.in/eS2YbZgC)
Later down the line we will need to change the model architecture. Token-by-token prediction is slow and error-prone. Language Diffusion models are already showing impressive speed gains because they can unmask multiple tokens per step and have the ability to fix token errors. Now do that in latent space, get rid of the tokenization bias, add variable compute per step and some form of hierarchical representations and voilà, you might get yourself the next era of models.
Act
You have a 3D bust and you want to approximate it as close as possible with a fixed amount of triangles.
Having 600 instead of 60 triangles makes a huge difference, but having 6m instead of 60k triangles doesn’t make a real difference other than being 100x more expensive to compute.
This is exactly the same of what we are seeing with LLMs nowadays. The bust is the data we are scraping from the web, and the triangles are the model weights. And we have the same problem: Increasing the model size by 100x will be 100x times more expensive but we will get improvements that are really hard to notice.
Why? The problem is inherently to the LLM structure.
Models spend the same amount of compute on filler words like “however”, “therefore” or variables when coding or doing math. No one cares which filler word to use, but if you even just flip a sign in a math formula it’s just wrong.
It’s funny how in the model has high entropy on filler words not because it doesn’t know, but because it’s not important which filler word to use (this point extends also to synonyms).
And if you don’t do good data curation, which is currently being neglected by academics and industry (the industry is slowly going more and more) than you will spend a ton of compute during training on things like filler words.
Where do we go from here?
First big step is data curation. Our current datasets are so massive that we rarely do more than a few epochs when training. So we get rid of the easy examples to shrink the dataset to give more compute to the hard samples. This might require something like curriculum learning during pretraining.
Many people thing we have LLMs because of the transformer architecture or some other model idea, but actually it has been data all along. First ImageNet, then web scraping, large scale human preference data, and finally data from verifier systems. You can have a LLM without the transformer architecture, but not without the data. (read https://lnkd.in/eS2YbZgC)
Later down the line we will need to change the model architecture. Token-by-token prediction is slow and error-prone. Language Diffusion models are already showing impressive speed gains because they can unmask multiple tokens per step and have the ability to fix token errors. Now do that in latent space, get rid of the tokenization bias, add variable compute per step and some form of hierarchical representations and voilà, you might get yourself the next era of models.
Act
express, not stress