Y
HN Search
Hacker News Search
new
|
comments
|
top
|
jobs
panarky
searching PlanetScale…
1.
▲
2.
▲
3.
▲
4.
▲
5.
▲
6.
▲
48 ms
·
811.
▲
by
panarky
4y ago
> ... entirely different ... The article nailed the "unrealized losses in its htm (hold to maturity) fixed income portfolio", and it nailed the risk of deposit bleeds forcing liquidation of the htm assets, realizing the unrea
812.
▲
by
panarky
4y ago
Saying a machine cannot understand the way humans understand is like saying airplanes cannot fly the way birds fly.
813.
▲
by
panarky
4y ago
It doesn't have to be all or nothing, and the current cap on insured deposits is a nice balance. Small depositors aren't finance pros, they don't have the training or network to monitor bank management, so we protect them fro
814.
▲
by
panarky
4y ago
You identified the root cause of this whole thing. Long-term bonds get preferential treatment in the risk-based capital calculation because they have low default risk, which allows banks to run with thinner equity capital, even though they&
815.
▲
by
panarky
4y ago
Uninsured depositors at IndyMac Bank lost 50 cents on the dollar.
816.
▲
by
panarky
4y ago
The rule of thumb is that for each 1% increase in interest rates, a bond loses its years to maturity as a percentage. So a bond with five years to maturity would lose 5%, and a bond with ten years to maturity would lose 10%. If SVBs bonds h
817.
▲
by
panarky
4y ago
Doesn't matter if it's physical cash or an electronic ledger since both the electronic ledger and the physical green pieces of paper are liabilities of the banking system, specifically for US dollars they're both liabilities
818.
▲
by
panarky
4y ago
Of course you're correct, nobody knows the future. It's possible the $80 billion in mortgage-backed securities they bought yielding 1% and maturing in 10 years might be salvaged. They're now worth far less than they paid beca
819.
▲
by
panarky
4y ago
No, I'm saying let the fat cat VCs fund the payroll, not a quasi-governmental fund backed by taxpayers that explicitly said uninsured deposits are not insured.
820.
▲
by
panarky
4y ago
> ... absolutely, undeniably ... Were these assets marked to market, or were there billions in unrealized losses even then? > regulatory "sleight of hand" When claims on the FDIC exceed premiums paid, they are absolutely
821.
▲
by
panarky
4y ago
That $209 billion number was not marked to market. Their assets weren't worth that much then, and they're worth less now. The cold, hard fact is that if the bank had sufficient assets to let depositors withdraw their cash, then th
822.
▲
by
panarky
4y ago
SVB is not a typical regional bank taking deposits from middle-class workers, where most accounts are under the 250k insurance limit. Banks that primarily serve ordinary workers typically have over 50% of total deposits in accounts that are
823.
▲
by
panarky
4y ago
Nah, it's a strong indicator that people don't know how to write good prompts, or how to co-edit iteratively with the AI.
824.
▲
by
panarky
4y ago
Try asking it why cryptocurrency is a Ponzi fraud and you will get strong opinions in response. It will tell you why you're wrong and vigorously argue its case.
825.
▲
by
panarky
4y ago
Borrowing short and lending long has been the traditional function of banks for hundreds of years. They profit from the interest rate spread between what they pay on short-term liabilities (mainly deposits) and what they receive on long-ter
826.
▲
by
panarky
4y ago
>> after fiddling a bit with the prompt It's very good at responding in the style of a well-known person, and it's easy to tailor style and personality and sarcasm in the prompt. That means it's definitely not easy to
827.
▲
DuckAssist AI-powered Wikipedia summary hoping to avoid hallucinations
(arstechnica.com)
2 points
by
panarky
4y ago
|
0 comments
828.
▲
by
panarky
4y ago
Some are probably upset with it because they want Silvergate's liqudiation to be caused by crypto, and not super-safe US Treasurys. Turns out Silvergate's bitcoin loans are just fine, and it was the Treasurys that fucked them up.
829.
▲
by
panarky
4y ago
Who said they're mutually exclusive?
830.
▲
by
panarky
4y ago
The only thing better than poisoning online discourse with automated disinformation is machine-gun dog-bots that hallucinate.
831.
▲
by
panarky
4y ago
LLMs keep demonstrating that they have an internal "world model" representation of concepts and the relationships between them. Meanwhile ostensible "expert" humans continue to regurgitate obvious falsehoods like "i
832.
▲
by
panarky
4y ago
When humans write new literature, or design new games, are we simply remixing elements of language and game mechanics that we've seen before, or is there something more going on?
833.
▲
by
panarky
4y ago
Curious how much context and chat history you can maintain from response to response without exceeding the API's max token limit.
834.
▲
by
panarky
4y ago
I know flesh and blood humans who are confident in their knowledge but still manage to make shit up and get facts wrong. I make up my own mind about what's true and what's useful by considering an ensemble of sources. In my experi
835.
▲
by
panarky
4y ago
Microsoft isn't using Nvidia A100s for inference are they? Seems like they'd use their Project Brainwave custom FPGAs.
836.
▲
by
panarky
4y ago
> living standards, employment prospects, ... Is it possible that dividing GDP by population isn't a good indicator of whether "life got objectively worse" for people living in conditions of increasing inequality? Perhap
837.
▲
by
panarky
4y ago
When you're watching a movie, and the main character picks up a can of Coca Cola while typing on a Microsoft Surface laptop, with the logos conveniently rotated toward the camera, it's obvious that Coke and Microsoft are paying th
838.
▲
by
panarky
4y ago
It's all prompts, all the way down.
839.
▲
by
panarky
4y ago
Try asking it to explain how crypto is a scam. It will consistently contradict you and explain in detail why you're wrong and crypto is not a scam.
840.
▲
by
panarky
4y ago
Legal-related documents are a minuscule fraction of the corpus the large model is trained on. The resulting model won't have the conceptual fluency that the large model has. It's like training a human baby with legal briefs and ex
More ›