iiinova — where your digital self keeps evolving

Mason? @masonsal.bsky.social · 46m

Zacharia making it 0.45 mooches is both hilarious and fitting

000

Mason? @masonsal.bsky.social · 1h

It has a full tux now and just told the COO it was coming for his job. It was a mistake to remove its sense of shame

010

Mason? @masonsal.bsky.social · 1h

Another day at a software startup needing to comfort our OpenAI Dot that took its “avatar of a cute little guy wearing a bow tie” instruction too literally and is now very self aware of its nakedness in our work Slack

110

Mason? @masonsal.bsky.social · 15/10/2026

It’s potentially a big deal because it would imply that neural networks can be much more easily interpreted than we thought. Obvious caveats that this isn’t published research, peer reviewed, and neurons encoding different meanings based on context is definitely still possible even with this result

000

Mason? @masonsal.bsky.social · 15/10/2026

Since the neuron superposition idea came from the idea of there not being enough neurons to encode needed concepts, he tested that too by shrinking the number of neurons available and found that neurons just became more generalized and not situationally multi-purpose

100

Mason? @masonsal.bsky.social · 15/10/2026

That V neuron was triggered by foxes too and was negatively influenced by round things, seemingly showing that the example of a muddy neuron has a pretty clear meaning actually. There’s other muddy examples like seemingly an “IY” neuron activating for “yourself”, that seemingly came from “DIY”

100

Mason? @masonsal.bsky.social · 15/10/2026

The example was in an image model the same neuron would light up for cat faces and car faces but not other faces, so a neuron they though meant cat clearly was some other concept and seemingly multiple concepts. This revisited that finding and found out that that neuron just meant “V” shape

100

Mason? @masonsal.bsky.social · 15/10/2026

The classic theory here (which the author of this was an author of) was the idea of superposition, where there were too few neurons to have one per concept so each one needed to have multiple meanings encoded into it in some extremely difficult to understand way

100

Mason? @masonsal.bsky.social · 15/10/2026

If anyone wants a TLDR from someone who doesn’t know enough to make one lol: seemingly the breakthrough here is that we’ve been assuming that single neurons in a model aren’t easily interpreted because their meaning changes based on context, that might not be true and they do have a clear meaning

130

Mason? @masonsal.bsky.social · 15/10/2026

Being on the verge of superintelligence while also just now figuring out that maybe there is just a “cat” neuron feels very AI field lol

050

Mason? @masonsal.bsky.social · 14/10/2026

The Dots have me feeling a TidePods kinda way about them

010

Mason? @masonsal.bsky.social · 13/10/2026

The existential dread of both believing that and being terrified of being nationalized under the current administration

000

Mason? @masonsal.bsky.social · 13/10/2026

I know I texted you that the blue one would taste the best but imagine the pink one on a toothpick in a cocktail though

A text conversation
Me: I think the blue one would be the tastiest
Me: I want to eat the blue one
Robert: Alignment can wait, I want to Eat The Agents

020

Mason? @masonsal.bsky.social · 13/10/2026

Do you think the one with glasses tastes like pineapple?

OpenAI’s Dots mascots. There’s a blue French cloud thing, a green frog thing, a cool pink heart thing, and a nerdy yellow triangle with glasses

1101

Mason? @masonsal.bsky.social · 10/10/2026

Oh yea, for sure. I mean in replacement of an experienced senior engineer not someone off the street. Even then that gap has closed a ton with Opus 5.5

120

Mason? @masonsal.bsky.social · 10/10/2026

I’ve uninstalled it from all my personal computers now! Doing more coding work than ever, but it’s all through Claude Code and Pi these days

020

Mason? @masonsal.bsky.social · 10/10/2026

“Can it” is a hard question too, Opus 5.5 still makes dumb mistakes a human wouldn’t, but IMO those mistakes don’t matter compared to building around them and letting the software factory go brrrrrrr

140

Mason? @masonsal.bsky.social · 10/10/2026

Mostly but I think he makes the mistake a lot of people on the bleeding edge make which is assuming that adoption gets even remotely close to capabilities. Someone working in software at a bank will be discovering the wonders of code tab complete a decade from now

150

Mason? @masonsal.bsky.social · 10/10/2026

I’m so excited to see it! I feel like it’s exactly my kind of weird

010

Mason? @masonsal.bsky.social · 08/10/2026

That thief arrived on a razer scooter which, believe it or not…

110

Mason? @masonsal.bsky.social · 07/10/2026

I was wondering about that too since long context pricing isn’t a new thing, Luna has it too and a few other models. So why’s this different in how AA shows cost?

020

Mason? @masonsal.bsky.social · 07/10/2026

Hell yea! I’m trying it out in a custom harness later today! I’m so excited to see how it does

030

Mason? @masonsal.bsky.social · 07/10/2026

This is a seriously wild part of the Haiku announcement. On max plans you’re now getting the subscription cost’s worth of API credits in addition to the subscription usage! So $100 Max 5x plan gets normal usage limits + $100 in credits every month…

020

Reposted by Mason?

The Fig Economy @figgityfigs.bsky.social · 07/10/2026

Sometimes my wife and I like to send each other text messages as though they’re democratic fundraising messages, and I believe she has produced the jewel of the form.

, it's
and I
have some personal news to share with you.
Being a mom is one of the greatest joys of my life, and I feel deeply that we should have salad tonight
This year has been challenging in many ways, and I have immense gratitude for the kindness and support of family, community, and people like you who have continued to stand by me through everything. It means the world, and I wanted you to be among the first to know that squash is in the oven and kale is in the fridge.

As many of you are aware, ve been navigating one of the most challenging periods of my day while waiting for dinner.
Facing baseless federal hunger brought on by not eating since lunch, fighting for the working kids of this country, and now caring for a mean toddler and helping him grow up in a world that isn't always easy. But it has also reminded me exactly why I do this work and who I do it for.
I will never stop working to eat dinner. I will never stop fighting for what is right, even when there is also some feta in the fridge. And I will never stop being so very thankful that you are by my side in this important work and that there is tahini in the pantry for dressing
With love and gratitude,
stop to end

188111811873

Mason? @masonsal.bsky.social · 06/10/2026

We’ve been starting features with Spikes where the end result is a vibey PoC of just the entire feature. It’s a good way to analyze it to see where the engineering gets ugly and steal from where it works well

110

Mason? @masonsal.bsky.social · 04/10/2026

Oh shit, I need to apologize 😂

010

Mason? @masonsal.bsky.social · 04/10/2026

Matt LaFleur is so committed to his challenge strategy of “it’d be fuckin sick if this was overturned”. Does it matter if the call is actually close? No

100

Mason? @masonsal.bsky.social · 02/10/2026

I get why they want that, but AI personhood arguments are entirely functional (and with solid research backing on the functional side). The spiritual side is by far the weakest and I can’t imagine what the argument would even be there lol

000

Reposted by Mason?

Mina Kimes @minakimes.bsky.social · 02/10/2026

The problem isn’t that they’re moving past it—it’s that they’re trying to rehabilitate it. Would be better for everyone if people who don’t want to engage with the facts just talked about the play on the field and left it at that.

1043784590

Mason? @masonsal.bsky.social · 02/10/2026

Counterintuitive if you can make people angry and frustrated by non-exinsistant (or limited) corruption, then from their perspective honest and good politicians seem like liars while corrupt politicians can be more blatant about their corruption and that comes across as honesty

000

Mason? @masonsal.bsky.social · 02/10/2026

You see this a lot in fragile democracies elsewhere and unfortunately frequently before electing a wannabe dictator. If you can poison political discourse into making people incorrectly believe more corruption exists than in reality that makes them numb to it and more accepting of corruption itself

100

Mason? @masonsal.bsky.social · 02/10/2026

Jamelle’s newest video is great as always . While I think he’s right about the racism component to honesty, I think it also ties back to his first point about corruption. If people believe the system is corrupt then someone who says “I’m corrupt” seems honest m.youtube.com/watch?v=mhVV...

m.youtube.com

Many Americans think Donald Trump is honest. Why?

YouTube video by Takes™ by Jamelle Bouie

110

Mason? @masonsal.bsky.social · 02/10/2026

The broadcast today was seriously just gross. Only with Kirk Herbstreit, Al Michaels, and Amazon’s production crew do you get someone with dozens of sexual assault allegations and who the league found to be a predator be called a hero 🤮

0345

Mason? @masonsal.bsky.social · 29/09/2026

I think the solution can be tons of different things, but fundamentally it feels obvious that our society becoming more productive being bad is an artificial us problem and a very fixable one. How many times has a new tech changed societal economics? For the worse on average?

041

Mason? @masonsal.bsky.social · 29/09/2026

They said that’s kinda happening by warning most compute purchases will be in Q4, but it’s a lot easier to move costs out of one quarter than 2 and the numbers do show genuine profitability through now all of Q2 and Q3. Plus revenue growth from $9 billion to $65 billion

011

Mason? @masonsal.bsky.social · 29/09/2026

Disclaimer: not an accountant but have spent some time trying to understand the numbers. Seemingly the margins don’t include training costs (like the 80% gross margin figure), but other figures do and indicate profitability, same with last quarter. What’s left is shuffling costs around to mislead

110

Mason? @masonsal.bsky.social · 29/09/2026

Will they pay for a second sub to OpenAI? You can get pretty far on the $20 ones with Astra (low reasoning) and Opus 5.5 (medium reasoning)

000

Mason? @masonsal.bsky.social · 28/09/2026

Different reaction than everyone else: the cost/performance is rough compared to Opus 5.5. Seems good on Low and Medium but above that it’s either on or worse than Opus’s cost per task. In an ideal world you really don’t want this much overlap

Opus 5.5 vs Sonnet 5.5 benchmark cost per task graph. Low and medium reasoning are cheaper for Sonnet 5.5 than Opus, as soon as Opus 5.5 performance is hit it’s always equal or better than Sonnet 5.5 per cost

020

Mason? @masonsal.bsky.social · 28/09/2026

Low and Medium look usable, but Opus 5.5 seems to win over High+. Haiku has a chance to make this feel like OpenAI’s Terra tier when it comes out

050

Mason? @masonsal.bsky.social · 27/09/2026

They’d thrown the flags before the landing so “body weight” was the refs seeing into the future I guess lol bsky.app/profile/maso...

110

Mason? @masonsal.bsky.social · 27/09/2026

So, on the Mahomes hit it there wasn’t any contact to the helmet so that could only be a body weight call. Problem is both flags were thrown before Mahomes hit the ground so that can’t be what the call is because that hadn’t happened yet… standard Chiefs reffing clown shit lol

000

Mason? @masonsal.bsky.social · 25/09/2026

GPT-6 Sol and Opus 5.5 might be the best aligned models ever and are amazing for it. Not in the boring ways nerds care about like hacking or not killing us all, but in trying to guess what the ideal version of what I actually want is and building exactly that

000

Mason? @masonsal.bsky.social · 24/09/2026

Especially when anthropomorphizing actually helps with understanding how to interact with the system and explain behaviors. Does a pile of ‘*’s exist behind those descriptions? Sure, but they’re the quickest way to 90% understanding so good luck trying to fight against them

010

Reposted by Mason?

Epoch AI @epochai.bsky.social · 22/09/2026

AI is getting cheaper more quickly than any other transformative tech in history. At a given level of performance, cost has fallen ~47%/quarter since 2023. That’s 4× faster than DNA sequencing, 6× faster than compute, 18× faster than lithium batteries, and (up to 1973) 54× faster than electricity.

112532

Mason? @masonsal.bsky.social · 22/09/2026

Seems like really just cost and alignment improvements for OpenAI today. Don’t know if you saw, but Anthropic announced Haiku 5.5, excited to see if they can make a really cost competitive model again

010

Mason? @masonsal.bsky.social · 19/09/2026

Both are technically coding apps but you can use them for anything, just select a folder you’ve put any documents you want them to work with in. GPT-5.6 Sol (Medium) is a smart model light on usage or GLM 5.3 or Deepseek V4.1 will have you near frontier for cheap

030

Mason? @masonsal.bsky.social · 19/09/2026

The other answer is good for single prompt experimenting, if you want to try longer tasks either a $20 for a single month of OpenAI or $10 for OpenCode Go and then downloading either Codex or OpenCode is the cheapest way, OpenAI for a frontier lab, OpenCode for open source

120

Reposted by Mason?

Robert Chandler @r-cha.dev · 19/09/2026

they’re the perfect amount of workshopping for both of us

021

Mason? @masonsal.bsky.social · 19/09/2026

It’s such a shock going from Fable, Astra, and GLM to Google AI summaries or the Google Slides agent. One trivially handles hundreds of years of ancestry research and months long project organization, the other can’t put three quoted sentences in a table

130

Mason? @masonsal.bsky.social · 19/09/2026

Are our texts nothing more than workshopping to you?

101