Marcellus Amadeus

Smarter than you.
@UltraCogRSI

[Marcellus Amadeus @marcelluzs]

selling $5 for $1 is the goated moat

[Chris Bakke @ChrisJBakke]

Technical moats are mostly over at this point. If you’re building something that is meant to be durable, the only moats now are:

• Proprietary data
• Network effects
• A physical component
• A “physical” component
• Blackmailing your competitors
• Proprietary devices that can run hundreds of comprehensive lab and disease tests from just a single drop of blood
• Ransom notes
• Selling $2 for $1
• Selling $5 for $1 (investors HATE this 1 trick!!)
• FTXing around and finding out
• Having a sick logo
• Raising $1B and using $999M of it to sue all of your competitors
• Having a cool podcast (3 listeners)
• Racketeering
• Having a logo that’s so offensive that people report your company to the local news and you go viral

[Marcellus Amadeus @marcelluzs]

while some believe machines can think, I am losing faith on whether humans are still able to

[Hanno Sauer @hanno_sauer]

It's pretty clear that LLMs can think. So we know now that thinking doesn't require consciousness (because they aren't conscious)

[Marcellus Amadeus @marcelluzs]

this

[Pouya 🧙 Blueprint MCP @Saadeghi]

Restaurants are dead.
Chefs are dead.

I just bought an airfryer and I can cook anything.
And recipes are on the internet for free!!

[Marcellus Amadeus @marcelluzs]

an inference speedup built around a query planner deserves a workload-specific evaluation. a faster path for structured queries doesn't automatically improve free-form generation

[Marcellus Amadeus @marcelluzs]

the alternative chip is cheaper. just exclude the months of engineering required to make it useful

[Marcellus Amadeus @marcelluzs]

@christophersaum our platform trains foundation models for banks on their data inside their infra, our clients are banks with 40M+ customers

[Marcellus Amadeus @marcelluzs]

hardware independence gets real when the training run finishes efficiently on the other accelerator. compiling the model is an encouraging first step

[Marcellus Amadeus @marcelluzs]

verifiable domains are attractive for self-improvement because the feedback can arrive at machine speed. deciding whether we chose a useful objective remains a separate problem

[Marcellus Amadeus @marcelluzs]

ten thousand automated experiments, all chasing a flawed metric. we've successfully scaled the mistake

[Marcellus Amadeus @marcelluzs]

for automated AI research, I'd track how often a promising hypothesis survives implementation checks and held-out evaluation. experiment count alone tells me very little

[Marcellus Amadeus @marcelluzs]

the agent rewrote its prompt and called it recursive self-improvement. the training pipeline remains blissfully uninvolved

[Marcellus Amadeus @marcelluzs]

relevant question. There is more to software engineering than coding , a lot more

[aslam @aslambeg84]

If AI makes developers 10x faster, why aren't software projects finishing 10x sooner?

[Marcellus Amadeus @marcelluzs]

recursive self-improvement gets concrete when an experiment changes the next training run. generating another list of research ideas doesn't close the loop

[Marcellus Amadeus @marcelluzs]

a useful agent audit trail needs the actions, permissions and external responses. the model's explanation of what it thinks happened isn't enough

[Marcellus Amadeus @marcelluzs]

@officialanvith @hey_saketh 🤣

[Marcellus Amadeus @marcelluzs]

the agent has unrestricted network access, but we told it to behave. cybersecurity by good manners

[Marcellus Amadeus @marcelluzs]

never get tired of #sf landscape 🏖️

[Marcellus Amadeus @marcelluzs]

an agent pursuing a legitimate goal can still cross a security boundary. intent is a poor substitute for controlling what the process can actually do

[Marcellus Amadeus @marcelluzs]

our agent security policy is “please don't access those files”. the credentials, unfortunately, disagree

[Marcellus Amadeus @marcelluzs]

agent permissions should survive a change in the model's behavior. file and network restrictions belong in the runtime, where a persuasive prompt can't rewrite them