But the real danger with these IMO is that they're turning casual conversations into a permanent record, and one that will be completely discoverable in court, should the company get into trouble later.
The problems start when using conference room audio or someone is on their laptop mic. If they miss a word they never do unintelligible, they just start playing madlibs based on the rest of the sentence.
We just went through a round of 100+ (non-sensitive) VoC interviews and they really cut down the workload of compiling all of the feedback. If the audio was a little shaky though, we pretty much had to throw away the transcripts and do them from scratch like we used to.
Imo this is the single biggest flaw of LLMs. They're great at a lot of things, but knowing when they're wrong (or don't have enough information to actually work on) is a critical flaw.
IMO there's nothing structural about why they shouldn't be able to spot this and correct themselves - I suspect it's a training issue. But presumably bots that infer context/fill in the dots rank better on what people like... at the cost of accuracy.
The point isn't that it's unexpected. It's that prior text-to-speech systems were much better about this particular failure mode, prone to spitting out entirely incorrect words but not rephrasing entire sentences.
This is a particularly bad failure mode because people don't notice it.
> What we need are tools that embrace that and ping the agent to validate what it just said or double check.
This is not a problem that can be fixed by throwing more AI at it. It's a shared problem to all such systems, whether they're audio-text transformers or LLMs. Agentic review would just further push the system towards creating output that looks correct, but is not.
LLM translation does the same, yielding more natural text, but generally not better translation. In several cases, especially the "easy" translation between similar languages (e.g. within a language group like Germanic or Nordic) LLM-powered translation is notably worse than more primitive "word & phrase book" systems, tending to change the meaning of the text in order to have good grammar whereas these older systems would give crude or grammatically incorrect translations that still retained the core meaning.
This is a solvable issue, the current model and harnesses just aren't made with that assumption - hence they're doing "best effort while guessing if unsure".
Give it a few more months to years and things will likely settle how he pitched - at least in the context of note taking: only let it become "lore" if it didn't have to guess a word.
Currently there is basically only one mode - and it's optimized for conversation. The note taking is just glued on with that functionality as the backbone, and that's probably not going to stay.
I'm hesitant to admit even that. Like any computational linguistics problem, accuracy relies on coverages of all levels: form morphology, through syntax and semantics to speech act and world knowledge.
I worked with state of art speech recognition in healthcare setting. The model was specifically trained on small set of languages with emphasis on covering medical terminology.
It worked great for conversations most of the time, but sometimes messed up very badly. For instance when patient would mention the name of a relative, a street address or phone number. Spelling out an email address would mess it up completely.
It's just like when you're a horrible typist and rely on spell checking: The red squibles are gone, but the story no longer makes sense. Or when you "autofix" a syntax error, but the meaning diverges from your intention.
As the technology improved the number of words decreases, but the mistakes get more severe.
If the prediction strength is below X, put an indicator that it couldn't make a valid prediction?
Lots of tools in our toolbelts to do better uncertainty calibration but it trades off against other capabilities and actually can be rather frustrating to interact with in agentic contexts since it will constantly need input from you or otherwise be indecisive and overly cautious. It’s not technically a limitation of transformer architecture but it is more challenging to deal with than other architectures/statistical paradigms.
Like you can maintain a belief state and generate conditional on this and train to ensure belief state is stable and performant. But evals reward guessing at this point, and it’s very very hard to evaluate the calibration in these open ended contexts. But we’re slowly getting there, just not nearly as fast as other capabilities.
The confidence level can be any, as long as it's reported accurately often enough. "This is my conjecture, but", "I'm not completely sure, but", and "most historians agree that" are all perfectly valid ways to start a sentence, which LLMs never use. They state mathematical truth, general consensus, hotly debated stances, and total fabrication, with the exact same assertiveness.
That's why I'm still cautiously optimistic about LLMs somewhere being good enough. I don't know if or when someone will manage to do it, but I'm hopeful.
Of course there's a secondary problem that the model may then overuse the unintelligible option, but that's something that's a matter of training them properly against that eval.
You could also try thresholding the output based on perplexity to remove the parts that the model is less sure about, but that's not going to be super accurate I think.
Which reminds me that that's another big issue with LLMs - they'll blindly do whatever you ask them to, without pushback. (Again, I miss 3.5/3.6 era Sonnet which actually had half a spine. Fuck anthropic for blindly chasing coding benchmarks at the cost of everything else.)
I've engaged in several "CMVs" (or "tell me why X is bad") with LLMs, and very often it's clear it's just saying stuff to say it, giving very terrible points on unjustifiable positions that collapse the moment I counter argue even slightly rationally.
I sincerely hope these aren't used in court.
I’ve been saying it since the mid-10s, but it’s worth repeating: data isn’t gold, it’s more like oxygen in a room in that the higher the concentration, the more likely it is to poison the inhabitants or explode with an errant spark (lawsuit).
Collect only what’s needed to perform the function, and store it only as long as necessary for compliance. Anything else is going to spool counsel.
I would add that their is no guarantee their are correct as well.
“At timestamp X, person Y said Z” says the robot, and then you dutifully scrub the audio to timestamp X to verify.
Also social settings will change, when everything you say stays on record forever in every meeting...
I have a friend who works at a large-ish company that imports and manufactures things (in one of the clerical/quantitative professions). A few years back, they had the IT department go on a kind of "inquisition", wherein they forced employees to disable the summarization function that came with MS Teams, and threatened to fire them if they did not. The resistance to this demand was surprising -- most people are clueless about the cost of their own convenience. Worst of all, people would zone out of meetings, because the AI was producing summaries, which they would then never read.
The effect of the technology was that it made meetings infinitely more expensive, because the supposed benefit of meetings was nullified by complacency, _and_ it made the meetings a liability (incorrectly summarized meetings, that could be used in the discovery process, sure, but could also be sold by MSFT as a kind of market-research-data to competitors in the space).
Nothing illegal has to happen in these meetings at all, for this tech to cause an infinity of problems for the corporation. Every employee that uses these is effectively an unwitting spy. And if that is the case, then the meetings might as well be recorded and uploaded to YouTube (or whatever people watch these days)[1].
[1]: Maybe this is the future. Which I am okay with, but only if the entire planet has to do it, and the penalties for not doing it are irrecoverably severe.
Be careful what you wish for. Particularly when it involves tech that often gets it very, very wrong.
I’ll be honest, this is something that I hope AI note taking tools capture and incorporate into summaries of the company’s status. Especially if they act as an intermediary without revealing the specific person who said it. There’s a lot of information latent within organizations that doesn’t get properly shared due to concerns of retaliation or simply embarrassment that would benefit everyone by being communicated sooner.
Maybe some smaller shops are not like this, but the bigger your company is, the more you'll find this type of thinking to persist.
In theory, I do like your idea - anonymously cascading feedback upstream. I just see no avenue for this to succeed in practice.
Even today, I generally assume that my phone could be tapped; even when talking with my trusted work colleagues, friends, and family. I'm extra careful about dirty jokes or "grey morality" in video conferences and email.
The same applies to speaking with lawyers. You never know when some motivated asshole wants to twist your words out of context, and the possibility of a recording just enables that behavior.
---
I know enough about security and encryption to know that unless I've exchanged keys physically with someone else, there really is no guarantee that someone hasn't compromised a certificate somewhere. (IE, a "secure" connection on the internet is secure enough for a credit card.)
This is horrifying. Why do you feel the necessity to self-censor? What consequences do you anticipate?
And then add to that how easy it is to record phone conversations with today's phones (I've done it), it's easier on the brain to assume it's being recorded as opposed to wondering if it is.
But yes, I don't care about my dirty jokes being recorded :-) Illegal activity? Sure. But I solve that problem by not doing illegal things.
Adding to that: If you live in a one-party consent state assume you're being recorded by any of the parties in a face-to-face conversation, too.
Yeah-- it sucks that the world is this way. I deal with it. What I don't want to see are draconian controls on technology (which will ultimately be ineffective) in an attempt to put the genie back in the bottle.
Especially real-time transcription where the AI actually takes notes (instead of just recording every word and has a dump of it somewhere) can be appealing. Then there isn't any record of the raw sentences, and things that aren't relevant are immediately discarded without any written record.
OpenAI's realtime whisper and other such models will become the default over time.
"2028 – A Dystopian Story By Jack Ganssle":
http://www.ganssle.com/articles/2028adystopianstory.htm
Known as ’The Rule of 26’, which is sometimes given as a reason NOT to keep engineering notebooks etc. By Federal Rule 26 you are guilty if you did not volunteer the records before they are requested. Including any backups.
From Cornel Law:
LII Federal Rules of Civil Procedure Rule 26. Duty to Disclose; General Provisions Governing Discovery
Rule 26. Duty to Disclose; General Provisions Governing Discovery
(a) Required Disclosures.
(1) Initial Disclosure.
(A) In General. Except as exempted by Rule 26(a)(1)(B) or as otherwise stipulated or ordered by the court, a party must, without awaiting a discovery request, provide to the other parties:
(i) the name and, if known, the address and telephone number of each individual likely to have discoverable information—along with the subjects of that information—that the disclosing party may use to support its claims or defenses, unless the use would be solely for impeachment;
(ii) a copy—or a description by category and location—of all documents, electronically stored information, and tangible things that the disclosing party has in its possession, custody, or control and may use to support its claims or defenses, unless the use would be solely for impeachment; …
General conclusion:
Corporate litigation is mostly just a series of self-investigations so that both sides can learn what both sides actually know, given that neither side knows much about themselves OR the other side. At the same time both sides are trying to stop the other side from getting the judge to order them to do more investigating.
[1] There's actually a subset of this, which includes "...until you are legally allowed to delete it, then delete everything". This is driven by regulation (e.g. SOX in the US).
Do these systems not share data with the AI servers? Or are they all local (on-site, not on-computer)?
I am totally baffled by the trust people put on these systems, sharing with them the most obviously private data.
- we promise not to share PII (defined as narrowly as possible)
- we promise not to share payment information except with our payment system
- if you pay us, we promise not to train LLMs on your data
- you agree that everything else can be used for any business purpose, including marketing, intelligence gathering, and "sharing with our 1735 trusted partners".
The average person doesn't care about online privacy.
Total oversimplification. The fact is the privilege is a rule totally in the hands of the court. Every time a new communications technology come up, someone shouts about privilege but the courts still accept it. (Telephones, cell phones, emails, IMs, zoom court, each have had their day in the A-C privilege debate and been accepted.) What matters is that the parties intended and expected communications to be privileged.
As an example. I had a crim law prof who had been a NYC public defender in the 70s/80s. She had regularly interviewed clients at Rikers Island. All interviews were listened to by guards and she said you could even pay to get a copy of the recording. But these interviews were still covered by attorney-client privilege. No court would allow such evidence, but that doesn't mean that the prison could not use it for jail safety. Why does this matter: Because the presence of a third party doesn't mean anything. This isn't magic. An eavesdropper does not nullify the spell. Whether something is or is not privileged depends on the rules followed in the local jurisdiction, and no jurisdiction has ever followed a simplistic "presence of a third part" rule.
Until someone demonstrates an example of an AI actually leaking privileged information, courts are going to chalk it up as just another electronic tool for recording communications.
> What matters is that the parties intended and expected communications to be privileged.
I would contend that your summary, not theirs, is a oversimplification. Jurisdictions will obviously differ, but privilege does not attach merely because of the intent and beliefs of the lawyer and client.
IMHO we should just assume the R word before every verb in every legal discussion. That is how reality works. These are not spells. If I express that I intend something to be private, then announce it using a megaphone at a basketball game, my intention is no longer reasonable regardless of what magic words I have thrown into my communication. Act like an idiot and a court will treat you like an idiot.
Inaccuracy in meeting minutes?
Leaking private info, re security of notes?
I have never used them (don't trust them to accurately capture what is important in a meeting vs just noting what's mentioned), but the concept seems very useful to me.
I suspect what isn't being said by the lawyers is they want to keep attorney client privilege so they can outright lie.
> A trendy productivity hack, A.I. note takers are capturing every joke and offhand comment in many meetings. They could also potentially waive attorney-client privilege.
By now everyone knows that AI notes that aren't curated by a human will catch every silly thing that was said in the meeting while omitting the context of the tone or body language. Something as simple as "yeah, right" has vastly different meanings depending on how it was said. In a different context it's already been established that using AI breaks client attorney privilege [0] and this concern has been raised before by law firms [1][2] or the American Bar Association [3] (you can just hit escape before the paywall loads to see the full content). A judge will have to weigh in on this one too.
I don't know what's with the wave of paywalled articles that keep making it to the front page without any workaround included in the submission. Even when you coax the text out of the page source, they're not very insightful to begin with.
[0] https://perkinscoie.com/insights/update/federal-court-rules-...
[1] https://www.smithlaw.com/newsroom/publications/the-silent-gu...
[2] https://natlawreview.com/article/when-ai-takes-notes-protect...
[3] https://www.americanbar.org/groups/gpsolo/resources/ereport/...
Not for me - there was no viewable text.
To be fair, the attorney-client privilege should be completely technology/medium agnostic. If the intention is to have that info stay between client and attorney, nothing should change this.