I find the money stuff newsletter by Matt Levine (bloomberg) great for this, the link is behind a paywal, but the newsletter is free. strong rec. todays newseltter https://www.bloomberg.com/opinion/newsletters/2026-03-11/pri...

From that newseltter:

> At the Financial Times, Jill Shah and Eric Platt report:

>JPMorgan Chase ... informed private credit lenders that it had marked down the value of certain loans in their portfolios, which serve as the collateral the funds use to borrow from the bank, according to people familiar with the matter. >...

>The loans that have been devalued are to software companies, which are seen as particularly vulnerable to the onset of AI. ...

From what i can tell the problem isn't that an individual who had cash to invest in a private (tech in this case) company goes down

the problem is that a company "private credit firms run retail-focused funds (“business development companies” or BDCs)" which took out a bunch of loans to invest in private tech companies is now having the underlying assets that they got those loans against (long term investments in private tech companies) valued lower.

the link im missing is what happens when people who also invested in BDCs want their money back, where their actual money is locked up in long term investments made to private tech companies, and their ability to get loans is now valued lower. I think this is called a "run" where if someone starts pulling money out, and ultimately you cant, then its a race to get your money out before others do, which applies to both the individuals and the institutional loans.

Note: my quotes are from the bloomberg newsletter i mention, which helped me, not the OP article. And i am writing as much to clarify my own thinking as from a place of understanding. I welcome clarification.

thank you so much! i suffered with this, and now i never will again!

ok, even that "few thousand examples" heuristic is useful. the usecase would be to run this task over id say somewhere in the order of magnitude of 100k extractions in a run, batched not real time, and we'd be interested in (and already do) reruns regularly with minor tweaks to the extracted blob (1-10 simple fields, nothing complex).

My interest in fine tuning at all is based on an adjacent interest in self hosting small models, although i tested this on aws bedrock for ease of comparison, so my hope is that given we are self hosting, then fine tuning and hosting our tuned model shouldn't be terribly difficult, at least compared to managed finetuning solutions on cloud providers which im generally wary of. Happy for those assumptions to be challenged.

Only to prompt thought on this exact question, im interested in answers:

I just ran a benchmark against haiku of a very simple document classification task that at the moment we farm out to haiku in parallel. very naive same prompt system via same api AWS bedrock, and can see that the a few of the 4b models are pretty good match, and could be easily run locally or just for cheap via a hosted provider. The "how much data and how much improvement" is a question i dont have a good intuition for anymore. I dont even have an order of magnitude guess on those two axis.

Heres raw numbers to spark discussion:

| Model | DocType% | Year% | Subject% | In $/MTok |

|---------------|----------|-------|----------|-----------|

| llama-70b -----| 83 | 98 | 96 | $0.72 |

| gpt-oss-20b --| 83 | 97 | 92 | $0.07 |

| ministral-14b -| 84 | 100 | 90 | $0.20 |

| gemma-4b ----| 75 | 93 | 91 | $0.04 |

| glm-flash-30b -| 83 | 93 | 90 | $0.07 |

| llama-1b ------| 47 | 90 | 58 | $0.10 |

percents are doc type (categorical), year, and subject name match against haiku. just uses the first 4 pages.

in the old world where these were my own in house models, id be interested in seeing if i could uplift those nubmers with traingin, but i haven't done that with the new LLMs in a while. keen to get even a finger to the air if possible.

Can easily generate tens of thousands of examples.

Might try myself, but always keen for an opinion.

_edit for table formatting_

I was aware of the patent, and agree i think its overly narrow and you could get around it easily. I think the reason we haven't seen it or something like it in another game (or i haven't but someone pleeeease id love to hear systems like it), is less because its not useful, or maybe its not useful as a plug and play because the only reason it works is because of the super exhaustive care taken on tuning its parameters and giving it enough variety to make it interesting to play.

Kinda like the dialogue/story paths in something like hades, where IIRC they made a whole system to manage it, but the reality is that system only matters when the tree is suuuuuuuuper complex. or maybe it was disco elysium, or both ...