Zoya Bylinskii

I describe my career journey from my undergrad days in AI, grad school at MIT, industry research and GenAI evals jobs, and recently founding my own company substack.com/home/post/p-19… to give others the confidence to make their own career pivots towards greater personal meaning

@perceptualinsights" target="_blank" rel="nofollow noopener">substack.com/@perceptualins… for weekly "perceptual bites", or how knowledge from perceptual & cognitive science can be translated into product differentiators, or “what you don’t know you don’t know” about human perception :)

Why A/B evals can lead us astray: open.substack.com/pub/zoyathinks…

15 "Mundane" Ways I've Used #ChatGPT in My Family Life in the Last 6 Months open.substack.com/pub/zoyathinks…

ChatGPT knows me so well 😍

Funny how easy it is to tell that ChatGPT has been used for an email or resume by all the semi-regular bold formatting distributed throughout the text, where what is emphasized is not always semantically consistent. To the human authors: do consider if the bold is always needed.

Love this demonstration of how differently we experience and explore our world from even the first days of life.

Creativity is still in the creator's hands. I had a vision in my head that I could barely translate into words. I used different models to see which would best understand me. None fully did, but they took me down new alleys, like philosophical chats over wine with new friends...

I was thinking about our place in the era of AI #superintelligence. I wrote down my partial, fleeting yet connected thoughts. And then I decided to leave them as is, a poem of sorts, in free verse. Unpolished and raw like the thoughts themselves: @zoya.gavr/we-are-the-superintelligence-a-poem-written-in-free-verse-89eca7acb5e2" target="_blank" rel="nofollow noopener">medium.com/@zoya.gavr/we-…

I have quite a few friends who are new or expecting parents right now, so I wrote them a letter, reflecting on what I thought it would be most helpful for them to know. Decided to share it more broadly in case it helps others too: zoyathinks.substack.com/p/a-letter-to-… #parenting

"When Evals are Not Data Annotation", another in my series of blog posts about #GenAI #evals: medium.com/p/when-evals-a…

A reflection of today's reality: seeing utter disappointment in my kids' eyes after trying to talk to Furby (the artificial "intelligent" toy from my childhood) and realizing that unlike ChatGPT, Alexa, or Ok Google, it can not answer their questions. That it randomly farts was seen as a redeeming quality.

In my previous post I talked about the properties of an #evals job (i.e., the job of building human-powered evaluations for GenAI tools) that requires some of the same skills HCI & Psych researchers already have. I got some questions after that post about how to better align oneself with a career in evals, given a background in these fields. I thought it would help to dig deeper into how evals has evolved into a job, why it’s different than the jobs that came before it, and what it borrows methodology from, as a way of describing the influences that feed into it, and as a result, the diverse tools it helps to have to be good at it. @zoya.gavr/evals-the-inter-disciplinarians-dream-c73be0a79757" target="_blank" rel="nofollow noopener">medium.com/@zoya.gavr/eva…

For my HCI and psych friends who are thinking about career paths in industry, there are some pretty cool opportunities in evals for GenAI right now: @zoya.gavr/why-evals-is-the-hot-new-thing-for-hci-and-psych-researchers-to-get-into-1c69e93c6f71" target="_blank" rel="nofollow noopener">medium.com/@zoya.gavr/why…

The irony! Was reading "The Design of Everyday Things" by @NormanDon on the plane, only to get off and end up helping a visually impaired young woman navigate the ladies room at SFO. Near-missing multiple walls (we had to turn 5 corners!), garbage bins, stools (?!), and people walking in multiple directions at once, made me see the #accessibility challenges through her eyes. "I hate hate HATE public bathrooms" she mumbled under her breath as we meandered through. And this was after standing in a long line because no one in the line realized there were a bunch of empty stalls around a corner 🙈🤦‍♀️

I've completed another short #AI #scifi story: zoyathinks.com/blog/longestar… imagining a world decades from now where all text once written by humans has been digitized, #LLMs have rewritten all this text countless times over, and the original content is unrecoverable. A hunt for "the originals" ensues by a pair of Linguistic Archeologists who stumble upon "The Longest Human Artifact". A question asked on @lexfridman's podcast episode with @sundarpichai (x.com/lexfridman/sta…):"What’s the most life-changing, five-word sentence hiding in that haystack?” which was generated by @GeminiApp, served as inspiration to develop some of the ideas in this story.

What worked great using @OpenAI's #ChatGPT for #storywriting: ✅1. Find the word at the tip of my tongue (e.g., "complete my sentence: 'frills and ___ of this approach' ", "what's another way to say Hilariously ironic") ✅2. Find a common theme in a group of words (e.g., "what is a common word to describe Ketamine, Piracetam, Ritalin, Benzodiazepines") ✅3. Brainstorm a word for a non-existent concept: (e.g., "can we give a job name to people that predict the probability of an automated accident?", "what companies other than tesla are working on autonomous trucks [...] which of those are likely to work in siberia in the future?") ✅4. Provide justification/evidence/feedback for my ideas (e.g., "what about autonomous risk actuary", "can a future vehicle have an automatic means to reposition the weight it contains?") ✅5. Summarize content from multiple sources (e.g.,"what are common situations with buses or cars in the Himalayas that cause accidents?", "what's a high slope grade on mountain terrain in the himalayas") ✅6. Rephrase information using new context (e.g., "can you rephrase this quote to talk about mudslides?") ✅7. Help me organize information in a structured format (e.g., "I'm going to have 3 tiers of emergencies, class 1, class 2, class 3") ✅8. Find concepts that match my descriptions (e.g., "i want the uber bus to go along a fault line in california, where?") ✅9. Provide additional examples of concepts (e.g., "what are natural cognitive enhancing methods - meditation, breathing exercises, what else?") ✅10. Use context from previous queries to answer new queries without requiring additional explicit context (e.g., "is there a road in hawaii that goes along a cliff edge?" ->"do any of these highways lead to a beach?" ->"would the drop be on the right or left side of the vehicle?") One of the best surprises was when at the very end of this long (multi-day) conversation, without any additional context, I asked, and got exactly what I was looking for:

In writing my most recent short scifi story zoyathinks.com/blog/operators…, @OpenAI's #ChatGPT was my trusty side-kick, not to replace my creativity or for the storyline, but as my research assistant to make sure the situations, locations, and environmental conditions I described were credible, to fill my knowledge gaps about how vehicles work, and to retrieve the words stuck at the tip of my tongue by playing "fill in the blank" with me. I'm happy to share the full dialog I had with ChatGPT while developing this story: chatgpt.com/share/687ec2b9….

Many are still scared and averse to the use of AI for creative writing. For instance, one top magazine for scifi stories clarkesworldmagazine.com/submissions/ has a strongly worded statement about banning authors who are in any way assisted by such tools, because there is currently no easy way to verify originality. But if authors could submit their #ChatGPT conversation alongside their work, and if there were automated tools to "grade" how #AI was used (e.g., to provide background information and find suitable words or translations VS to write fragments of the final text), wouldn't some of these concerns be dispelled? And couldn't we use a similar approach for grading student work in #educational settings? It would have a similar flavor to submitting the cheatsheet you used for a test, or pre-registering your scientific experiment, or including your raw data as supplemental material with your paper. Couldn't the incentives be set up in a way that allowed AI tools to be used but required you to #preregister and then submit your conversation? If the contents of the conversation were not well aligned with the project in terms of timeline (start and end timestamps) or quantity of content (e.g., suspiciously little back-and-forth), then you could also potentially catch cheaters that submitted the "wrong" conversation. We could probably come up with a set of constraints and verification pipelines whereby submitting the wrong conversation (i.e., a fake conversation that omits critical pieces of the dialog, e.g., where AI helped to do more of the writing) would end up as being more work then just submitting the real, full conversation.