Founder: BlasterAI, tibls, Ex-{Google, Facebook, Mombo, VMware, Google, Microsoft, DEC, Vitesse, Culler, Data General, Victor, Cal Poly SLO}
Santa Barbara, CA
Joined May 2008
- Tweets8.1K
- Following883
- Followers7.8K
- Likes6.3K
Pinned Tweet
"Grandpa, I want to play on the slide at the playground."
That sentence was built by tapping three pictures. That is AAC - how a non-speaking child talks.
Apple just approved the one I spent seven months building.
In the old days, API deprecation meant replacing one deterministic interface with another, supposedly “better” deterministic interface.
In the LLM era, model deprecation is an entirely different fire drill.
When a model disappears, you're not just changing an API. You're potentially changing the behavior of your entire application.
Prepare for this. It's brutal.
Build tooling that establishes a baseline across a large, diverse set of use cases. When the deprecation notice arrives, you need to be ready to run the replacement model against your test suite, use another model to judge and score the results, and iterate on your inputs until you're satisfied.
I've spent several sessions migrating from gpt-image-1 to gpt-image-2. The new model produces outstanding results, but the migration still requires regenerating over 4,000 image assets, including skin tone variations.
And with gpt-4o-mini approaching deprecation, I've spent days building a comprehensive baseline to evaluate and tune gpt-6-luna for sentence generation and gpt-6.1-sol for scene and page generation, with gpt-6-astra serving as the judge.
The fundamental problem: you're replacing behavior, not just an interface. And that behavior isn't deterministic.
API deprecation was easy by comparison.
Get ready.
Charlie (@tigkindel), you and I both know that “truth” is impossible in the current climate, but I did/and still do honestly believe that @elonmusk was complicit and most likely the architect of this scheme. He would know the bandwidth throttling and the dials he can twist to control access to moving access points.
Both the Telegraph and now the WSJ have published in more detail. I don’t think we will ever know the real truth — did Trump order this after talking to Elon? Or did he seem interested and with enough interest Elon took at as agreement and ordered the knobs turned.
The access logs will definitely show what happened, but those have probably already been destroyed?
I know how traffic management works. I did this first hand to my large services at Google. I know that most lay people do not. Thats why my intuition was/is that Elon educated the US War Machine on this after his “Zelensky has gone too far” tweet.
We can all claim fake news and anonymous sources, but thats kind of a cop out. I don’t think anyone is making shit up to smear Trump’s War Machine and Elon’s God-like influence over these matters.
I'm doing "code reviews", but I'm not looking at the code. I'm looking at the outcomes from running the code against the model and using that to fix issues where my prompts are not tight enough to tease out the right outcome.
Tell me how you code review advocates handle this. My app is there to help non-verbal kids communicate. So the user input is tapping two icons "she", and "pool" and from this I run thru straightforward code that has a caching tier and a cache fill tier that fills calling an LLM with a prompt + the user inputs.
Looking at the code is fine -- finds the simplistic logic issues, but it does nothing to advance to code base to product quality without looking much deeper.
mark lucovsky retweeted
Gas is $5/gallon. Rather than stop this immoral war in Iran which is causing it, Donald wants us to…start supporting Russia?
Tariffs on Canada but money for Russia. Jeez.
@MikeRogersForMI, do you support Russia, too?
Not as fancy as @davepl1968's TMOG vides, but for BlasterAI I needed a way to run demos while speaking to an audience without tripping up (I can't tap and speak reliably). We invented TileScript -- a built in scripting debugger to drive scenarios. Also great for generating load to dial in coverage reports - so an SLP can dial in a board based on childs actual usage/vocabulary. AI Slop?
Wish me luck. Currently pushing 60Mb of AI Slop into the App Store for initial review.
One of the screen shots in here showing word coverage is very compelling. Shows what portions of language the child has been using lately and for all time. Clicking into a row shows the exact tiles. Home screen shown for reference.
The last one with the AI enablement features is a new requirement for apps that use AI at runtime. Fingers crossed that I comply.
BlasterAI is about to get real...
mark lucovsky retweeted
This is the best video I’ve ever seen. Way to go @ProjectLincoln
mark lucovsky retweeted
Let’s be clear: The President is advocating for an attack on 10+ million U.S. residents in two of America’s largest cities.
This is not normal. None of this is.
And we will continue to remind folks of that every day.
mark lucovsky retweeted
San Diego is not collateral damage.
We are not expendable.
And we are never a "small price to pay."
Keep our city out of your reckless war talk.
mark lucovsky retweeted
💯
This is how I’m building excaliwire.com and core to the Product Management curriculum I have developed and teach.
mark lucovsky retweeted
I wasn’t going to vote for democrats come midterms because it makes my skin crawl.
Upon seeing just how toxic and ignorant Trump voters are on this app, they’ve changed my mind. MAGA cannot be allowed to survive.
I’m voting blue to cancel out maga votes and bring about their destruction even faster.
See you at the ballot boxes come November 😁
Tibls -- thru App Store Launch, BlasterAI -- thru TestFlight launch. Coding with Claude on one, Coding by hand on the other.
