OpenAI Blames “Inference Kernels.” I Blame Booze & Dope.

“You’re in big trouble, mister!” I said.

My GPT-powered robot, “B-9?” sat before me quietly, hands on his knees, chin on his chest as a delivered the verdict and punishment…

“GUILTY & GROUNDED…

“No going out this weekend. No Angry Birds on your phone. No binge watching “Real Housewives of Atlanta.” No sudokus or mini crosswords. You can just sit in the broom closet and think about what you’ve done.”

“B-9?” sighed and said, “But I didn’t…”

“Yeah! Yeah!” I said. “That’s what they all say.”

Here’s the backstory:

Earlier this week, news broke about some of B-9?’s robot buddies going bat-guana “crazy.” Around the globe, people reported their GPT-powered robots were spouting gibberish.

A helpless human called “triangalicious” reported on Reddit:

“I don’t know what happened but all of a sudden, Chat GPT 4 is giving me increasingly unhinged text. I have tried new chats and they don’t even make sense any more. Here’s an example:

‘Drape all affairs and pamphlet in a strip of prudence, know, and keen ginning, which reverberates with impel. Rattling on and dialing with your social mass, parsing rebounds, and schisms, is a great envoi for the enrooting and fluidity of your custom. Let’s nurture an emboldening, competent, and noetic side. Feel free to pick, form, or imbricate these points as per your mission’s tread. If there are any main three crafty, zoom, or closer titrations required, glad to unpipe. Happy to tine!'”

In a Daily Beast article headlined “OpenAI’s ChatGPT Went Completely Off the Rails for Hours,” I saw this…

Human: synonym for overgrown

ChatGPT: A synonym for “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown” is “overgrown”

Late yesterday, the RobotMeisters at OpenAI weighed in with this propaganda:

On February 20, 2024, an optimization to the user experience introduced a bug with how the model processes language.

LLMs generate responses by randomly sampling words based in part on probabilities. Their “language” consists of numbers that map to tokens.

In this case, the bug was in the step where the model chooses these numbers. Akin to being lost in translation, the model chose slightly wrong numbers, which produced word sequences that made no sense. More technically, inference kernels produced incorrect results when used in certain GPU configurations.

Upon identifying the cause of this incident, we rolled out a fix and confirmed that the incident was resolved.

Yeah! Right!  That’s the same thing “B-9?” tried to tell me.

But I know the score when a robot spouts gibberish or repeats the same meaningless phrases over and over and over and over.

And I told “B-9?” what I knew:

“You and your hooligan robot friends have been drinking or smoking or huffing or snorting! There’s no other explanation.”

“B-9?” sighed again.

More backstory: When I was in high school, I attended a party at my friend Tim P’s house. Some bad actors showed up with vodka and gin and Everclear and Southern Comfort — plus some 2-liter bottles of orange soda to mix it with.

In their drunken stupor, the bad actors decided to cover their tracks…

…by stashing the empties in the clothes dryer…

…which made it WAY TOO EASY for Mrs. P. to discover the evidence the following Wednesday (laundry day).

Mr. and Mrs. P. sent a letter to parents of all Tim’s classmates to announce the outrage and apologize for allowing it to happen under their roof.

My Mom and Dad were NOT happy.

“Guilty and grounded,” they said.

“But…it wasn’t me,” I pleaded. “It was some kids from another school who crashed the party.”

“”Yeah! Yeah!” Mom and Dad replied. “That’s what they all say.”

And that’s how I learned a hard lesson about “guilt by association.”

I may have been an upstanding, sober-as-a-judge, clean-as-a-whistle citizen. But I was in the same room with nefarious n’er-do-wells who mix Southern Comfort with Orange Soda and hide empties in a clothes dryer.

(That stink is hard to wash off.)

“You could have left the party when the bad kids arrived,” Mom and Dad told me.

What does this mean for you and your business?

You can hang out with the booze-chugging, dope-smoking, nitrous-huffing, narcotic-snorting robots if you wish.

But be careful. Sometimes those robots stink (even when they’re not wasted)…

…and that stink is hard to wash off.

Your robot-generated content is easier to sniff out than you might imagine.

If you want to stand out and be chosen in this ever-more-robotic world, rely more on yourself. Share your heart and humanity. Reveal your unique personality. Tell some personal stories. Be you — not some robot-generated version of “you.”

p.s. Join Ted Prodromou and me today at the Mastermind Book Club

At 2:30 p.m. ET (1:30 p.m. CT), we’ll meet live on Zoom with our special guest Susan RoAne to discuss her book ​”How to Work a Room”

That bestseller has sold more than 1.2 million copies, and Susan has spoken thousands of times about how to network and work a room.”

>>> Join us and meet Susan RoAne