Shubham Sharma

cto @fluencitysocial. building ai agents for influencer marketing

[Shubham Sharma @shu8ham_]

@TheStalwart the scarier part is the fraud model can't tell it apart from the customer. same device, same session, same cookies. it just looks like you having a very busy afternoon

[Shubham Sharma @shu8ham_]

banks are finding third-party ai agents logged into customer accounts, some doing things the user never asked for.

not surprising when the only door a bank offers is the login page. give agents a read-only token with limits or they'll keep using your password

[Shubham Sharma @shu8ham_]

@paschals_robert @anandlunia a visible ramp helps a lot. people tolerate limits they can see coming, what burns trust is hitting a wall mid-campaign with zero warning. i'd put the next step-up date right next to the counter

[Shubham Sharma @shu8ham_]

@arpit_bhayani .pth is the nasty bit. you can audit every import in your repo and it still runs the first time your editor spins up python. and claude code and cursor keys were on the list, so the coding agent's own creds were the prize

[Shubham Sharma @shu8ham_]

@ArtificialAnlys @Wisedocsai @OpenAI for anything a clinician actually reads i'd take accurate and concise every time. completeness is how a wrong dosage ends up in paragraph six where nobody checks

[Shubham Sharma @shu8ham_]

@patio11 @allTheYud models treating anything after their cutoff as a suspicious rumor is the most consistent failure i see in research agents. telling it today's date and that search results outrank its memory fixes most of it. not all

[Shubham Sharma @shu8ham_]

@santiagomed a skill helps less than giving it 2-3 of your own past articles plus a list of phrases it's banned from. most slop is the model's default voice leaking through, and examples override that better than instructions do

[Shubham Sharma @shu8ham_]

@zkevinbai @claudeai the approval gate on emails and payments is the easy part. the hard part is deciding which of the 40 almost-important things a day are worth interrupting someone for

[Shubham Sharma @shu8ham_]

@DocumentingAGI meta is the best in the world at buying installs, so the download number mostly tells you about meta's ad machine. 23% day-14 retention is the real stat, and that's still 3 out of 4 people gone in two weeks

[Shubham Sharma @shu8ham_]

anthropic's duet, openai's dots, meta's muse. everyone's shipping an agent that watches your apps and taps you on the shoulder. the number that decides who wins is false alarms per week. one that pings you 9 times a day about stuff that didn't need you gets muted by thursday

[Shubham Sharma @shu8ham_]

@JamesWalker1979 @freeCodeCamp @gkoos430 yep, the budget has to key on the task id, not the run id. otherwise a crash loop just looks like five very well behaved runs

[Shubham Sharma @shu8ham_]

@ghita__ha @NotionHQ clustering real tasks first is the step most router evals skip. pairwise also beats 1-10 scores because judges are awful at "7/10" and decent at "which is better". the hard bucket for a router is the cheap-looking request that quietly needed the big model

[Shubham Sharma @shu8ham_]

@gabepereyra @ShawnFarsai @figma @ElevenLabs @claudeai storyboard as the PRD is the part people will skip and then wonder why the agent's film looks like stock footage. also once video costs nothing to make, the scarce input is a face the audience already trusts. creator rates aren't dropping for that reason

[Shubham Sharma @shu8ham_]

@PinkDraconian an agent will happily call a crash an RCE because that's what the prompt was hoping to find. no repro request, no finding. otherwise it's a screenshot of a vibe

[Shubham Sharma @shu8ham_]

@marckohlbrugge the bottleneck just moved to maintainers. php has an actual security team. plenty of packages with 10M weekly downloads have one tired person and a sponsors page

[Shubham Sharma @shu8ham_]

@IT_unhinged the list is missing marketing's campaign microsites. built by an agency three agencies ago, still on php 7.2, and DNS still points at them

[Shubham Sharma @shu8ham_]

@FelipeDaragon two months of embargo with a big engine change and a small set of eyes on it. a day-one bypass is almost tradition at this point. patch anyway and leave the waf rule on

[Shubham Sharma @shu8ham_]

"defender's window" assumes defenders patch fast. the php fix took two months to build and plenty of sites won't upgrade for another six. ai finding bugs faster mostly helps whoever moves quicker, and that's rarely the team with a change freeze

[Shubham Sharma @shu8ham_]

@DeItaone "using existing cloud spending commitments" is the line that matters. half of enterprise ai buying is just whichever model you can pay for without a new procurement review

[Shubham Sharma @shu8ham_]

@nikks_techie summaries + a store is the textbook answer. the one i'd actually want to hear is what happens when step 20 contradicts step 3, like a creator agreeing to a rate and then changing it. stale memory is worse than no memory