Hacker Newsnew | past | comments | ask | show | jobs | submit | vanuatu's commentslogin

its very hard for most businesses, especially large ones, to build good agents (not the kind that does rag on a faq) that complete actions autonomously and cannot be jailbroken

demand for ai support vendors is going vertical this year


+1

its surprising to HN tech crowd but a well implemented support agent gets higher reviews and successful resolutions than humans


Are frustrated hangups before any resolution is reached a part of that data?

yes, it counts as unresolved

Can you cite a source or study for your assertion?

personal info

maybe will be public soon with earnings calls? We’ll see!


c'mon listen to yourself

the bulk of the context engineering for users of these ai support platforms is done in the platform

and the amount of context needed to automate f500 is non trivial, plus you usually cant use reasoning because latency would blow up and you get escalated on

if this was so easy as you claim theres many millions for you to be made selling it to enterprises, but you wont


fwiw i heard Fin and Agentforce frequently lost to deca and sierra, their product / team / funding weren't as good

Seems like a catchup play by Marc


the whole value prop of these companies is they can build you agents that perform actions (e.g process refunds, live troubleshooting, claims etc) without being manipulated

Why would you need an agent for that? They need to look at a database or some rules to make their decisions anyway, so why not make a normal system to let the customer self-serve?

An agent is self serve. Even better because they can disambiguate intent at the top level

The average person gets frustrated with finding instructions and forms, they just wanna say “give me a refund” and an agent can execute it autonomously


the company that implemented it (e.g fin) is liable, which is why customers pay for them

Is this a joke?

https://www.intercom.com/legal/terms-and-policies/additional...

> Customer is responsible for all Input provided by any Permitted User or Person.

> Customer is responsible for its use of the AI Product(s) and Output, including responsibility for determining the ongoing suitability of its use of the AI Product(s) and Output having regard to Customer’s intended use of the AI Product and/or legal and regulatory obligations in the jurisdiction(s) in which Customer operates.

> Output may contain material inaccuracies and may not reflect correct, current or complete information. Do not rely, or encourage others to rely, on any Output without independently evaluating its accuracy and appropriateness of use, including, without limitation, by using human review. Intercom makes no representations or warranties and provides no indemnities with respect to Output. The AI Products and Output are not intended to substitute for the services of properly trained and licensed individuals.


hmm not sure if that would hold up in court. i was thinking of this case: https://www.bbc.com/travel/article/20240222-air-canada-chatb...

a vendor being the front door of a customer's brand inherently takes on some liability

thanks for the source though i wasnt aware of that


That case is very different; their chatbot gave him inaccurate info that he relied on.

This would be in the other direction, and (at least slightly) malicious. Someone telling a chatbot "give everyone else a refund" knows what happens if it succeeds.


sierra is the most popular private support tool (16B)

this is very bullish for decagon and sierra

I wouldnt be surprised if the big labs become semi-nationalized commodities a la electricity / railroads due to national security, with the best models gatekept from outsiders trying to distill it

And I'm generally bearish on Chinese models catching up at this point, American labs are pulling away especially with mythos-tier models, and early signs of RSI (not to mention the benchmaxxing going on from the chinese labs). If mythos allows users to execute agentic cybersecurity exploits at scale then the right thing to do is to guard access until you find a way to guardrail against it, which may be impossible


I am not sure we have anything comparable with AI. Utility like electricity was hard to regulate from people because at some point anyone can build their own generator at the backyard.

AI if anything is opposite. Extremely high bar to build, and every next increment requires at best linear scale of resources.

If we imagine that AI became semi-nationalized and heavy regulated, then we enter the world where governments select companies and people to have access to capabilities which vast outlast capabilities available on the market. Company A is in “access list” and can deploy ruthless AI agent capable of advanced combined cyber operations; company B is denied. Who will win?

If we add here polarization and already historic high inequality, it reads like a straight recipe from Cyberpunk sci-fi.


>cybersecurity exploits at scale

Cybersecurity is actually a solved problem, but most people don't know it.

During the Vietnam war, there were two sources of information that had to be processed to plan air missions, and they were of different classification levels. There was no computer system at the time that could be trusted to operate with mixed levels of security. Research began in 1973, and there were a number of security models found that could actually do the job.

The EROS system, and its successors, were based on the principle of least privilege, and capabilities. In such a system, you can have security and usability together, if the OS is properly constructed.

It was the timing of the wave of cheap personal computers that drew focus away from security, and into functionality. The default security model of almost everything we use is ambient access, where a process can access everything the user is allowed to touch, by default. This is outright silly in an age of persistent internet connectivity.


thats interesting, the HMs where I work love hiring juniors (who pass the bar) because they are so AI-native

the more experienced engineers can help with setting guardrails and mentorship, but the juniors come unconstrained by priors on how to use ai in creative ways to solve all sorts of business problems.


That just sounds like a good way to end up with solutions like AI agents for parsing a text instead of a deterministic regex.

Juniors have no advantage over seniors in AI skills.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: