Weekend reading: If the drugs do work

Comments on this entry are closed.

• 

  Maven

  1 Azamino June 20, 2026, 10:31 am

  The City AM article about over reliance on AI by lawyers is really a story about risk. My shop has paid for a significantly restricted version of ChatGPT (restricted by it being trained solely on our data). Our version is useful but limited when compared even to the free version that can be accessed by anyone online. We knew this when the decision was made to restrict the training and it is confirmed when we find / catch staff using the full version*, what we term ‘shadow IT’. When challenged a common defence is that the firm’s version is too watered down and lacks bite.

  * almost always while wfh, dimwits still email files back and forth, AirDrop from BYOD is what the smarter ones use.
• 

  Maven

  2 Ducknald Don June 20, 2026, 11:20 am

  >When challenged a common defence is that the firm’s version is too watered down and lacks bite.

  Is that the software or the staff member?
• 

  Maven

  3 Frugalist June 20, 2026, 12:13 pm

  @Azamino It does seem to be the direction of travel, at least for big corporates, to customise a GPT and put the appropriate guardrails in. As you say though, it’s all too common for staff to get frustrated… or simply feel that they get more out of the agent they use at home, and then you’re into blocking whack-a-mole territory. I don’t feel like anyone has ever quite solved the problem of shadow IT – as you say there’s always a loophole like airdrop somewhere in the chain. On paper it’s user awareness etc, but in practice… tricky.
• 

  Mogul

  4 ZXSpectrum48k June 20, 2026, 1:52 pm

  I think for any sensible corporation, understanding what data the AI model has been trained on is pretty critical. You may want it trained only on internal data; you clearly don’t want the generic ChatGPT getting to train on your data. We’ve introduced rules making it gross misconduct to use generic AI tools with our data. I don’t think that is over the top.

  More interesting is the sheer cost of the usage of AI, especially after price rises from the likes of Anthropic and OpenAi. My firm were heavily encouraging usage. So running up $1k/day is nothing special for a front-office user!

  I wonder how long that will last. Yes, token costs are falling but the tokens required to solve a problem reliably only ever seems to rise. I’m wondering if productivity is really rising and what the return on that money really is. I still think we’d be better off just going back toward using in-house LLMs better optimized to the issues we need to solve. I ran a model four times on Friday because it refused to accept that 6 and 7 were not the same number! Another $1k gone …
• 

  Mogul

  5 flotron June 20, 2026, 4:55 pm

  @ZX Ha! Sounds like the model was trolling you with the meme which was popular amongst the kids last year.
• 

  Mogul

  6 Delta Hedge June 20, 2026, 5:55 pm

  (#4/@ZX): Yep. See my #660 here:

  https://milestone-vector.live/weekend-reading-first-they-came-for-the-call-centres/#comment-1954211%3C/a%3E%3C/p%3E
• 

  Mogul

  7 Delta Hedge June 22, 2026, 12:32 pm

  Novo Nordisk gets written up lots by value crowd Substakers. Often mentioned in the same breath as PayPal (or even more dinosaur like Western Union) and the fallen SaaS darlings like Adobe, Constellation Software and Duolingo.

  Who knows?

  I detect some slight signs of bottoming out. Very fragile but always is such. Frank Herbert’s “the begging is a very delicate time” and all that.

  I tend now to think that avoiding value traps (for those inclined towards the art form of Grahamite stock picking) is to look at how many times the shares have turned over in the drawdown and whether or not there’s any institutions left on the register.

  Idea is that a company that’s turned over its share count 5x in the current drawdown is less risky (within some limit) than one that’s only done 3x, but more so than one with 15x share count turnover in the same period, because this signals seller exhaustion.

  Shares go up and down for all sorts of reasons but, fundamentally, it’s just supply and demand.

  Once the institutions, the mom and pop hold for lifers (not), the insiders, the indexers (because it dropped out of trackers due to reduced cap weight not meeting index inclusion rules), and the turnaround narrative tourists (and the deep value seekers) have all sold out, then there’s noone left to sell FAPP and (theory is) the only way then is up.

  I’m not sure myself.

  I think you also need a Klarmanite big margin of safety in Net Tangible Book Value to Price and a catalyst of sorts in “out underperforming” low expectations.