@Salesforce MVP | Founder @sfindiadreamin | I live and breathe Salesforce , CLDTO Asymbl Building the future of work today with Intelligence Investor, Speaker
Kochi & Dubai
Joined October 2010
- Tweets23.4K
- Following3.9K
- Followers6.7K
- Likes9.1K
Pinned Tweet
It's always amazing to meet the man who's vision that started this all ! Thank you @Benioff for everything you do for the @salesforce #ohana and spreading love around the world ! Thank you again for bringing #TrailheadDX to #India. #myHero #Visionary #CEODiaries #Pledge1Percent
Can AI search tell if a resume fits a job? Not in my tests. Both models did barely better than a coin flip.
For matching people to jobs, I'd use an LLM with a clear scoring rubric instead. Write-up: shivd.me/embed2-local-cloud-…
Should you mix keyword search into AI search? It depends on what you search.
In my tests it helped a little on contracts and hurt a lot on code.
Same mix, opposite results. Test it on your own documents: shivd.me/embed2-local-cloud-…
AI does three jobs in a data pipeline: designing it, running thousands of small tasks, and testing improvements. When one chat assistant does all three, it grades its own homework and mistakes reach production. Give each job its own model and its own check.
For architects building AI data pipelines: define the failure test before the first model call. A small hand-checked sample is enough. Without it the pipeline has no target to optimize against, and you learn about the error from a downstream report.
Monday list: write down your thin core, pick one decision and build its layer, write the failure test first, and track checking spend separately from answer spend.
Day 09: From December, ChatGPT and Codex, OpenAI's AI coding tool, share one usage limit.
Why care: one heavy coding week can leave the rest of your team short.
Do you know who watches that limit for your team?
My moves: shivd.me/pulse-09
#OperatorsPulse #PulseDay09
7/ JetBrains coding apps trusted eight stolen digital IDs, so someone on your network could fake a real server.
Why care: a normal system update won't fix it. The new JetBrains versions do.
Your move: ask your developers to update every JetBrains app.
blog.jetbrains.com/security/…
8/ What I keep coming back to: somebody set each of these tools up, but I can't tell who owns what they spend or reach after that.
Your move: put a name next to each one.
Share this with whoever pays your AI bill.
The whole issue: shivd.me/pulse-09
Shivanath (he/him) retweeted
Two AI models from different families checked the same 200 records. They agreed on 194, and none of those shared answers was wrong. Every error was in the 6 they disagreed on. So a person reviewed 6 records instead of 200. One test, not proof, but a cheap pattern to copy.
Two AI models from different families checked the same 200 records. They agreed on 194, and none of those shared answers was wrong. Every error was in the 6 they disagreed on. So a person reviewed 6 records instead of 200. One test, not proof, but a cheap pattern to copy.
A pipeline tried four date formats when the file used a fifth. 99.4% of dates came out empty and the row counts still matched. Passing the row count is not the same as keeping the meaning.
One team spent $545 checking AI answers and $128 producing them, on a 203,317-case data cleanup. Producing an answer is close to free now. Trusting it is the expensive part.
In staffing, "candidate" means skills and availability to a recruiter, a pay rate and start date to payroll, and right-to-work and consent to compliance. Same person, three objects. Your core should hold the person, not all three.