barbacoa

I don't know if this is the correct way but it is the pattern I have fallen into.

Written code is the new compiled code. You never look at it, not even once. Documentation is the new written code.

Every repo has a /docs folder that holds the full specification of the code. Full architectural block diagrams, edge case behavior, style, problem statement, etc. I closely review these docs and make sure that the agent follows these specs and ensures the specs are updated after every code change. In theory an agent could fully rebuild the system from scratch using this documentation.

Maybe this is just pearl clutching, but it helps for when people ask architecture questions and I can answer and pretend I know what's going on.

>Theranos generated around 7.8 million test results for Arizona consumers alone. Of those, roughly 834,000 had to be voided or corrected. These were real people getting medical information they relied on.

Those are rookie numbers.

When Merck hid data showing Vioxx was causing heart attacks and death, not only did no one go to jail, but the fines were less than the revenue of the drug.

Vioxx would go on to kill an estimated 50,000 people before being pulled off the market.

Holmes made the mistake of having too few lobbyists and FDA insiders.

It's not obvious that the EU has the innovation ability or startup culture necessary to compete with spacex. For example in the whole of the EU, Spotify is the only company less than 25 years old and with a market cap over $50b.

On the flip side it stops me constantly for hacking reasons when I tell it to make/install custom firmware to bypass hardware restrictions and get root access on certain systems. I have to constantly remind the agent I work at the company at question and have their permission.

What's ironic is that Apple does all their mechanical CAD work for their products in Siemens NX running in Windows VMs.

>But that is considered success is it not?

Not if you took out a $50k loan for an early gen solar system and it now only saves $30 a month.

Berkley is defiantly not representative to most of the USA. They have strange regional politics that restricts most residential construction.

>250sqm, 30cm thick brick for load bearing + perimeter walls, 30cm rockwool continuous external insulation, aluminium frame triple pane windows, glazured roof tiles, floor heating + wood stove backup, 15kwp solar panels installation, deep well backup, etc.

Even in Mississippi what you describe would not qualify as a "top of the line house". In fact that would be pretty average working class home.

The kind of homes a lawyer or doctor could afford to buy in low cost area like mississippi can easily be 2 or 3x that size and resemble small castles. Think elevators and wrap around double staircases.

Be careful with these kind of long term sell back agreements.

In the USA lots of early adopters of solar did so under net metering agreements where excess electricity is sold back at full price. After a number of years it was costing the utility companies too much money. So they updated the terms such that residents with solar buy electricity at retail prices (~$0.25) but sell it back at wholesale (~$0.04).

They are saying that AMD's new Epyc Venice CPU has 16 memory channels allowing up to 1.6Tb/s of bandwidth. Which is higher bandwidth than most non-HBM GPUs.

So full CPU local AI inference may become viable option in coming years.

My new favorite new passtime with frontier LLMs, keep telling them "I just ate a [non-food object]". Eventually it gets stuck in a loop of telling me to call 911, unlock my front door, and lay on the floor in case I lose consciousness.

I'm running qwen 3.6 27b at 8bit quantization and 262k context. It takes 53gb of vram on my system.

The connection does not look standard so probably need to find some kind of OEM interposer card, which would allow you to connect some slimSAS cables that in turn connect to pcie adapter cards.

You'd need at least threadripper cpu/motherboard to handle that many pcie slots/lanes.

Not to mention that a baseboard like that is not going to work with a standard ATX power supply, so you'd have to provide your own power solution.

If that doesn't sound fun to you, you're probably better off buying Claude tokens.

This is from the USA site:

https://www.ebay.com/itm/157742745616

It's always a gamble buying used electric but per the description, fully decked out server with 256gb vram.

Though at 2U it is going to sound like a 747 taking off from your office.

4x A100s with the baseboard for $9k (USD).

https://www.ebay.com/itm/336632412718

Buy one and let me use it when you're sleeping.

>sticker price is 7899.

It's not even worth it at that point.

You can get a used enterprise grade SXM baseboard with 4-8 V100/A100 GPUs off eBay at a similar price. That will even get you actual HMB ram and NVlink. Along with 10x the AI performance, assuming you don't care about your electricity bill of course.

Try looking into Ryzen AI Max 395. AMD made a CPU/GPU soc with unified memory specifically for ai inference. Can buy mini PCs with up to 128gb ram.

America innovates, China replicates, Europe regulates.

I applied for a founding engineer with a startup though a "who's hiring" post. Had 3-4 rounds of interviews, met founders, did a project with them. I had competing offers and ultimately didn't work out. But it was a positive experience.

Apple briefly was selling Mac studio with 512 GB of unified ram, meaning all that was available as vram.

They are likely including xAI and the massive training costs that entails in order to make it sound like all their ventures are losing money.

Since Twitter is now part of SpaceX, does that mean Twitter is classified as a common carrier too? Can Twitter no longer refuse to allow certain individuals use their platform?

Honest question here.

Apple wants to vertically integrate. Their AI strategy until recently was to develop their own LLM models that were small enough to run on device. But massive scaling is what makes LLMs so powerful, so all their internal models were terrible and unusable.

Basically they bet that compute efficient LLMs were the future. That bet was wrong and the opposite came true.

>I believe in science.

I pray to science every day to deliver us from this heretical reign of RFK jr.

It's the new pattern you see from this kind of reporting. Innocuous beginning, omit the middle context, then mentioning the end reaction.

"Young mathmatics professor buys cabin in Montana, then gets arrested by FBI!"

"Austrian born artist wants to make Germany better, then the UK firebombs Dresden!"