The good, (and yes, I did try it)
Astra is an incredibly strong model and it just recently landed on the $20 plan within days. Some of the main highlights is that it includes computer use, and long running tasks it can actually finish. If you're a builder reading this, the ceiling on what one person can automate expanded once again this month alone, meaning that there are more use cases for automation, and a more heightened worry for jobs being automated.
I am not here to deny that. I am also not going to tell you it is what the launch week said it was. I tried it. It is overhyped. I sit somewhere in the middle on most of these, where I will not wave it off as marketing and I will not repeat the press release either, but TBH, that middle is getting harder to stand in.
So take the capability seriously and take the framing around it with a grain of salt.
The bad is the speed
Everybody argued about whether or not Astra was "revolutionary" or "shit" (Reddit seems to think it sucks). While I love myself a good argument about whether a product sucks or not, no one really stopped to think about the speed at which this was developed.
Early on, a unprecedented jump took the better part of a year. Claude 3 to Claude 4 was about 444 days. Claude 4 to Sonnet 5 was about 404. Then look at what happened inside the fours: Opus 4.5 in November, 4.6 in February, 4.7 in April, 4.8 in May. Three months, then two, then one. Fable 5 to Fable 5.1 took about three months, and 5.1 landed eight days before Astra did.
AI is getting faster and faster at development. Every time a new model releases, it is more powerful and all encompassing, and it takes only a fraction of the time it used to.
This is not my own inference at all. Today, Anthropic's own Alignment Science Lead named recursive self-improvement as happening faster than they thought, in the same breath as saying they have no plan for superintelligence alignment and are not clearly on track to get one. He put the odds that AI kills every human at over 10% inside a decade. He still works there.
Artificial general intelligence means a machine that is about as good as a person at everything that is not physical. We are not there. But the thing that scares me is not the capability of it, it is that the people closest to it are describing something THEY don't even have full control over, and shipping on it anyway.
The ugly is the marketing, and it works on you
That being said, I've come to find that these companies tend to overstate what a release can do. People get outraged about the overstatement. The outrage puts the product in front of everyone who was not paying attention, and a good chunk of them go and try it themselves, because that is what people do when everyone is talking about something.
It is the same exact mechanism that sold matcha and Dubai chocolate in every coffee shop in North America. It is how trends move and it is how startups scale. There is nothing exotic about it. What is new is that the product being sold this way is also the tool now writing most of the copy and a growing share of the posts arguing about it.
Go on Instagram tonight. Half of what you see is a partnership and the tell is not subtle any more, it is just constant. There's so many popular videos of UGC creators talking about some fictitious thing that happened and how the AI tool that is sponsoring them fixed all of their problems. Go look at their videos. Hell, even go look at streamers. Neon is doing a 24/7 live stream with Higgsfield and Astra. These companies are literally paying influencers to promote their AI slop.
So what can we even believe
The question I keep landing on is: what is even real and organic anymore? Can we trust anything anymore?
Forming your own opinion used to be the easiest part of thinking back then. You read a few things, you weighed them, you decided. Now, we're in a bit of a pickle. When reading articles and consuming content that is often meant to either drive an agenda or promote a business paying tens of thousands of dollars to create a UGC network to market their stuff, sorting a real claim from a generated one takes way more work that most people do not have literally any time for. Judgment is turning into a skill, and skills that are scarce become commodities. That is the real cost of generative AI as it is being used right now, and it lands well before anything anybody is calling a risk.
Where we stand
Regulation is going to have to catch some of this, and the EU is further along than anyone else right now. I would rather that arrived as a rule than as a disaster.
We build with generative AI. It is most of what we do. We also turn down work, and the line we hold is that we will not take a project whose whole point is to make something harder to tell apart from the real thing. That is not a position that costs us much yet. Ask me again in a year.
