In the video at the end you can see it’s the former. They’re drawing a digital representation of the commands the emulator is sending to the “sewing machine”.
As a user of a lot of coding tokens I’m most interested in latency - these numbers are presumably for heavily batched workloads. I dearly wish Claude had a cerebras endpoint.
I’m sure I’d use more tokens because I’d get more revs, but I don’t think token usage would increase linearly with speed: I need time to think about what I want to and what’s happened or is proposed. But I feel like I would be able to stay in flow state if the responses were faster, and that’s super appealing.
To quote McEnroe, commentating Wimbledon this year: "Father Time, undefeated." Djokovic is mentioned in the article and has only just ended his dominant era, and is still ranked 4th in the world at 38. So we did get some very long runs in there, and I would imagine just 3 years ago or so people would have expected some mid to late 20s/early 30s guys like Zverev or Fritz to be having their turn. Both of whom, some asterisks.
Instead we got this young duo / lightning in a bottle situation; and I expect that both Sinner and Alcaraz are likely to be playing dominantly into their mid 30s barring injury, or maybe Alcaraz buying a nightclub in Ibiza and retiring.
The cli subscription actually actively cannibalizes the API business in my experience. I think this is a product decision: if you use it to code, they want to control the user experience.
If you use it to back up 100,000 MAUs, then they want you to use the API.
I was originally an API user but the cli subscription is so much cheaper that I switched over. This is a combination of th CLI getting much more useful and reasoning models using many more tokens.
This is nicely written, and the style is fresh after a year of reading LLM writing whenever I click.
Interesting to me is the audience - this article is appropriate for a middle schooler on up to read, and assumes no technical knowledge at all, or very little. Documentation targets and the social culture around them have varied since I started reading documents like these, but this sets an especially low bar for reader knowledge.
I like it. On the one hand, I could have gotten much more technical detail out of a version that used industry-specific words and was half the length, but on the other, I got what i needed, and learned much about the author to boot.
It also impressed me, as I'm not sure I'd have that sort of dogged patience. Samba is one of those incredibly useful pieces of open source software that sometimes I feel I take for granted.
I won't take it for granted now.
As a side note, the documentation has been pretty darn good too. I set up an AD server in Samba just from the docs, with a bit of additional help from Stack Overflow. It was only after I had finished that I determined that I could do what I needed with just the basic Samba user/groups. (My needs were not complicated enough to justify the extra overhead of AD.)
Samba was arguably the way Linux snuck into corporations in the late 90s. Microsoft’s server offerings were truly terrible in that era, compared to getting a very functional Unix-a-like for free out of commodity hardware. Terrible and expensive. The issue was, you needed to be able to do group file sharing to windows desktops, full stop. IT teams wanted Linux, bad, for nerd and quality of life reasons. They also wanted Apache, not IIS. As Samba stabilized, Linux boxen quietly started appearing without notice or fanfare, and never left.
2003 was a completely different era for technical detail. That kind of detail normally only sat in corporations.
People would jump in, contribute their part, write a bit and leave. Folk were willing to contribute ideas to your project and assist in parts you got stuck. You learnt and understood by studying the source you were given or obtained. There were less expectations, if you had a half-baked thing folk gave you grave .
Code was more optimistic, fun and free for all. You didn't get lynched for not having a license or for not using $LANG. You grabbed something from Sourceforge and ran with it.
Why are you not using Python!? Why are you coding in TCL. PHP, pathetic. eww, you use IRC? lol perl.
Forums were rampant, the internet was a friendly place; LAN parties were awesome and the internet had rainbow coloured fences not grey greased walls like now were climbing over are a struggle requiring you to leave with one or the other. The world wasn't as depressed and a new phone was a new phone and not just a rehashed Android UI.
Where did we go wrong is an answer that cannot be answered other than we seem to keep making the same mistakes over and over again blaming whoever.
Granted I was 15, so naivety. Thirty Seven this year and still waiting for someone to code the time machine to relive those days.
I agree with everything you said, apart from "That kind of detail normally only sat in corporations.". I do not agree with it, because it applied to many open source projects.
I do, I never understood it then, I do now. Hybrid Ircd is what I used for my server. You were cool if you used Ratbox, and Unreal I never unstood why you needed +a and +q
Documentation was never fun so I'll use my age as an alibi. I was just a script kiddie and recalling downloading MB's of scripts and uploading them on to polarhome [0] where some awesome swede was hosting shell access to different systems on his home DSL. Me kitted with my three-hour limited 56k and my fathers work 128K ISDN line.
Oh well, nostalgia over back to work, writing IT architecture docs.
I’m 50, and I can tell you exactly what happened - AOL. Plus the death of the BBS. Usenet started changing radically in the late 90s, and the independence BBS operators had wasn’t appealing enough to keep them running dial up services, nor was it generally appealing enough to keep people logging in over modem when the baby internet was available.
All that said, I want to be clear that I’m still partial to TCL - an elegant highly usable language, and happy to fight you over it.
Driving character out of the software world. Spotify couldn't whip the llama's ass even if it wanted to. And it doesn't want to. Spotify is a boring office worker that we're professionally cordial with only because they run the stockroom.
If you looked at 'old' software companies that existed at 2000, they were boring companies like IBM. Really what most people don't want to admit is most of the software world grew up and got old.
Re Jeff and Sanjay - they recently were on Dwarkesh together I believe - so it looks like the partnership is still going strong. Regarding Dean over Ghemawat facts, the vibe from the convo is that Sanjay is the (very slightly) junior partner of the two, or at least he lets Jeff do more of the talking. Very, very nice vibes hearing them talk, and their war stories are clearly nuts.
The one thing I noticed when I worked near Jeff and Sanjay and talked to them over coffee is that Jeff is the smart one, but Sanjay is the wise one.
Jeff always had an idea how to make something a bit faster using a clever trick, but Sanjay would respond by coming up with how to express the abstraction in a way that other mortals could comprehend, or just telling Jeff it wasn't a good idea because it would make things more difficult to maintain.
Jeff was also prone to dad jokes, Sanjay's humor was far more subtle. Both were awesome to talk to and one of my proudest moments was when Jeff read a document proposal I wrote ("Google should get involved in Genomics Research and Drug Discovery") and took it seriously.
At the time I first got involved, Google Health was still a thing but it was clear it was not going to be successful. I felt that Google's ML (even early on, they had tons of ML, just most of it wasn't known externally) was going to be very useful for genomics and drug discovery.
Verily was its own thing that was unrelated to my push in Research. I think Larry Page knew Andy Conrad and told him he could do what he wanted (which led to Verily focusing on medical devices, which is a terrible industry to be in). They've pivoted a few times without much real success. My hope is that Alphabet sheds Verily (they've been trying) or just admit it's a failure and shut it down. It was just never run with the right philosophy.
Calico... that came out of Larry and Art Levinson- I guess Larry thought Art knew the secret to living forever and by giving him billions Art would come up with the solution to immortality and Larry would have first access to it. But they were ultra-secretive and tried to have the best of both worlds- full access to Google3 and borg, but without Googlers having any access to calico. That, combined with a number of other things, have led Calico to just be a quiet and not very interesting research group. I expect it to disband at some point.
Isomorphic is more recent than any of the stuff I was involved in, and is DeepMind (specifically Demis's) attempt to commercialize their work with AlphaFold. However, everybody in the field knows the strategy of 1. solve protein structure prediction 2. ??? 3. design profitable drugs and get them approved... is not a great strategy because protein structure determine has not ever been the rate limiting step to identifying targets and developing leads. I agree I don't really see a future for it but Demis has at least 10-20 years of runway before he has to take off or bail.
All of my suggestions were just for Google to do research with the community and publish it (especially the model code and weights, but also pipelines to prep data for learning) and turn a few of the ideas into products in Google Cloud(that's how Google Genomics was born... I was talking to Jeff, and he said "if we compress the genome enough, we can store it all on Flash, which would make search fast but cheap, and we'd have a useful product for genomics analysis companies"). IMHO Jeff's team substantially achieved their goals before the DeepMind stuff- DeepVariant was well-respected, but almost every person who worked on it and related systems got burned out and moved on.
What is success, anyway, in biotech? Is it making a drug that makes a lot of money? What if you do that, but it costs so much that people go bankrupt taking it? Or is the goal to make substantial improvements to the technology, potentially discovering key biological details that truly improve people's lives? Many would say that becoming a successful real estate ownership company is the real destination of any successful pharma/biotech.
Whoa. Finally someone I relate with! Thanks for such amazing intel!
In my opinion forays into biology by moonshot hopefuls fail for one of two reasons: either they completely ignore all the current wisdom from academia and industry, or they recruit the very academia people who are culturally responsible for the science rot we have at this time. Calico (and CZI, and im starting to fear, Arc) fell prey to the latter. Once you recruit one tenured professor IMO youre done. The level of tenure track trauma and academic rot they bring in can burn even a trillion dollars into dead-end initiatives.
IMO (after decades of daydreaming about this scenario), the only plausible way to recreate a Bell labs for Biology is to start something behind a single radical person, and recruit the smartest undergrads into that place directly. Ensure that they never become experts at just one thing so they have no allegiance to a method or field. And then let that hoarde loose on a single problem and see what comes out. For better or worse neuralink seems to be doing that right. Just wish they didnt abuse the monkeys that much!
To me success in biotechnology is if I measurably help make a drug that makes a person smile and breathe easy that would otherwise not have. Surprisingly easy and hard at the same time.
Having worked with them I would say Sanjay is certainly NOT the "junior" partner. Nor vice versa. They have different strengths but I couldn't say that one or the other is a better engineer overall.
I'm not a medical researcher, but I am a computer guy; I was struck by something very different in the papers - the abstracts at least refer to "AI CAD" as what they're testing - no software information, no versioning - on the CS side, this stuff is of paramount importance to make sure we know how the software performs.
On the medical side, we need statistically significant tests that physicians can know and rely on - this paper was likely obsolete when it was published, depending on what "AI CAD" means in practice.
I think this impedance mismatch between disciplines is pretty interesting; any thoughts from someone who understands the med side better?
The link is essentially a press release. The information you want is (sorta) in the actual paper it describes *.
"The images were analyzed using a commercially available AI-CAD system (Lunit INSIGHT MMG, version 1.1.7.0; Lunit Inc.), developed with deep convolutional neural networks and validated in multinational studies [1, 4]."
It's presumably a proprietary model, so you're not going to get a lot more information about it, but it's also one that's currently deployed in clinics, so...it's arguably a better comparison than a SOTA model some lab dumped on GitHub. I'd add that the post headline is also missing the point of the article: many of the missed cases can be detected with a different form of imaging. It's not really meant to be a model shoot-out style paper.
* Kim, J. Y., Kim, J. J., Lee, H. J., Hwangbo, L., Song, Y. S., Lee, J. W., Lee, N. K., Hong, S. B., & Kim, S. (2025). Added value of diffusion-weighted imaging in detecting breast cancer missed by artificial intelligence-based mammography. La Radiologia medica, 10.1007/s11547-025-02161-1. Advance online publication. https://doi.org/10.1007/s11547-025-02161
Regardless of title, it’s a good little reminder that I hadn’t thought of - skills can use skills. This makes sense - a skill is just a pre-loaded context Claude instance, so why not? But I also tend to think of one skill at a time. Thanks for the write up.
You can make a router skill that describes how to use the other skills together. I'm experimenting with this now but my core problem is still How to Make Claude Code Skills Activate Reliably.
During testing today I asked a task I knew should have activated a skill and claude just did it without the skill instead.
(I've rephrased the sentence by removing "just" -- the novel thought here is that the agent's use of a skill can automatically trigger another skill, which is somewhat emergent.)
> It’s also 100% vibe coded. I’ve never seen the code, and I never care to, which might give you pause. ‘Course, I’ve never looked at Beads either, and it’s 225k lines of Go code that tens of thousands of people are using every day. I just created it in October. If that makes you uncomfortable, get out now.
Followup request - digital emulated sewing machine! Should be a breeze based on the blog!
reply