Four years ago, my husband and I started writing online.
No product. No customers. No clue what we were building.
$200 articles for a grant-funded Web3 ecosystem site.
We knew it wasn't a sustainable business model.
We didn't know what we were good at, or what we should be doing instead.
We kept going anyway.
Year two, I cut my corporate software job to 80%.
I used the home-office hours to go all in on what we were building.
In the male-dominated field I worked in, this move was unusual. For a woman in that field, it was more so.
My husband went further. He's a doctor. He cut his practice to 20%.
We were each pulling two jobs. The corporate one. The other one.
A year later, it had a name: The Founder's Voice.
Founder-led communication for Web3 founders.
Year four. Full-time. Clients we chose. Calendars we own. And a one-and-a-half-year-old son.
This @thejustinwelsh line is the one I keep coming back to:
We're making more than the old salaries, on work that's ours.
That was the goal.
(Thanks Justin for naming it cleanly)
People underestimate how strong Google’s position in AI actually is, just like they did with Meta.
Google has the research depth to make fundamental advances, its own computing infrastructure, and products that billions of people already use.
I’d be surprised if Google doesn’t end up leading the AI race.
Across our projects @GoogleDeepMind, we are witnessing a genuine phase transition in the pace of scientific discovery. You will see a glimpse of this in our releases in the next few weeks.
There’s no standard for agent memory yet, and Anthropic’s post is upfront about that.
What it does have is a pattern that keeps repeating across very different tests:
1. Purpose >> structure: Rules about what to record made the strongest models worse. A clear purpose worked better.
2. Quality beats quantity: Every note the agent saves is something it may have to sort through later. A useful memory should save the agent more work than it creates.
3. The record is the foundation: Underneath the notes is the full log of what happened. Wrong notes got fixed by checking them against it.
4. Memory needs upkeep: Notes go stale and some are wrong from the start. A regular pass that checks them against the log did more than writing better notes in the moment.
None of these principles are tied to a particular model, even if how well a model uses memory clearly is. Which is why I’d keep memory separate from any one model or platform.
That’s the best way to make sure that the next model can build on the work you’ve already done.
Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. anthropic.com/research/compo…
I spent some time going through what @typesafeai has built with Jev. The design choices here are really interesting
1. It changes how the model answers
You define the possible answers, and Jev returns choices, scores and probabilities. It gives up generating free-form text and can answer independent questions in parallel
That’s a big part of the efficiency argument
2. It could make much more automation affordable
Businesses constantly need to sort requests, assess urgency and decide which requests need a person’s attention
If those checks become fast and cheap enough, you can afford to do them much more often
Across millions of requests or documents, that could change the economics of an entire product
3. There are still important limits
Jev won’t write a customer response or explain its reasoning. It can also choose the wrong answer, even when that answer follows the format perfectly
The headline gains come from specific benchmarks. You still need to test the accuracy, uncertainty estimates and costs on your own work, and decide what the system is allowed to do
I can see a lot of uses for this alongside models that handle the writing and more complex reasoning.
Choosing which model does which part of the work is going to become an important product decision
I got access to Jev earlier today (thank you @hackgoofer). I have run ~5,000 requests so far, (which cost me around $2!), across classification, model routing, intent, steering, and many other things.
tl;dr, Jev enables a new intelligent decision-making primitive, separate from deterministic code and LLM calls. This allows a class of decision-making that was neither suited to dumb, unintelligent code, nor to slow, expensive LLMs.
It is super fast and cheap, and I think I will likely end up making a few Jev calls to every LLM call I make in my product. I think probably any company using LLM requests today can probably add a Jev call pre and/or post LLM calls to quite literally make their product much better for free, and have better tool calling behavior in many cases.
I happened to have a personal benchmark for this as I’d been working on a ton of proactivity and classification tasks. I have been using the deepseek flash and more recently gpt 5.6 luna family of models as reasonably smart classifiers with low latency. Think questions like:
- Did this conversation output contradict something they’ve mentioned before?
- Should we send a followup message to this user based on our rules?
- It’s been a few seconds of silence. Should we proactively send a message?
In the past, I’ve been forced to write a bunch of what I call decision chains, mostly because an LLM classification call is very expensive in TIME (avg 4s), and less importantly can cost quite a bit if run on every message.
Imagine a normal chat app. If you added 4s to every response to figure out if the response is good before sending it out, that ends up being pretty bad. So instead, I usually have to write some code that is a crude heuristic that runs quickly and decides whether to run the classifier. Obviously this sucks because you call the classifier many times that you don’t want to, which makes your p95 bad, and you also miss cases with the heuristic, and you also have to manage all of these weird chains.
With Jev, it’s cheap enough, and fast enough (p50 ~150ms, p95 ~350ms in my testing!) that you can easily run it every turn. Heck you can reasonably run it before generation AND post generation, for any application that isn’t realtime voice, and still feel snappy.
But this is just one use case. Think: smarter model routing, better context packing, better responses, better observability for intent/tags/safety, smarter retries and so much more. By simply thinking about the inputs and outcomes you want to enforce, you can use Jev to supercharge most model calls and reduce bad user outcomes. The more “quirks” a model has, the more valuable it ends up being.
It’s a bit weird and unintuitive using Jev. Generally, you want to decrease the # of questions you ask a classifier, or it makes more mistakes. In fact, you might want to ask your questions kind of in a compound way, because the reasoning happens in a shared scratchpad of sorts. Adding questions muddies the scratchpad and makes it take longer.
With Jev, you feel incentivized to go the other way, to formulate your query as a set of independent questions. It doesn’t feel like adding more questions decreases your performance on others.
You can go a bit deeper to improve tool calls. Many model tools are things like turning on settings, or other things. You can easily improve models that are not very good at tool calling with Jev, by simply figuring out when to run them. You can do a pre-LLM call to figure out when to unfurl different tool definitions, in order to make your main LLM run better, you could run a background task with Jev + another LLM to reduce tool and context burden on your main LLM, and free it to be responsive.
I’ve only scratched the surface of my testing but very excited!
some crypto founders think switching ecosystems will magically bring adoption and solve all their problems
at some point, you have to stop lying to yourself and ask whether you and your team are the reason things aren't working
that’s when you actually have a chance to change something
I can see more people asking what happens to the data they share with AI and how much control they actually have over it. I think those questions are only going to get bigger as AI becomes more and more useful
We’re going to give these systems a lot of context about our lives and businesses. Over time, that could become some of the most useful knowledge we have. And I want people to have a real say in how it’s used
We’ve already made some of this possible with Walrus Memory. And I think there’s a lot more we can build around the same idea using @WalrusProtocol as a foundation
There is a super shady data control setting on everyone's ChatGPT which makes it seem like your data can be used for training even if you explicitly say to *NOT* improve the model for everyone (1/9)
The easiest way to understand the deal is to start with what @NVIDIAAI actually sells
NVIDIA makes money from the chips, networking and software used to train and run models
A future dominated by a few closed-model companies would not necessarily be bad for NVIDIA. Those companies already create enormous demand for its products
But it would leave much of that demand in the hands of a small number of extremely powerful customers. Many of them are also investing in their own chips and infrastructure
Open models create a very different market
When a model can be downloaded, adapted and run anywhere, far more companies can build AI for themselves. Startups can customize models for specific products. Enterprises can run them privately. Governments and universities can deploy them on their own infrastructure
That spreads AI activity across many more companies instead of concentrating it inside a few model providers
This is the future Jensen has been describing. Not one model doing everything, but many models being used for different jobs, at different costs, across every industry
But releasing open models is not enough. Developers still need a practical way to find them, understand them and turn them into products
That is where Hugging Face matters
It is often described as the GitHub of AI models, but it does much more than store model files
Developers use it to discover and compare models, read about how they were trained, test them in a browser, load them through the Transformers library and either use them through an API or run them on their own infrastructure
According to NVIDIA, more than 18 million developers and 200,000 companies already use the platform
Hugging Face removes much of the work between an open model being released and someone actually building a product with it
That is the important part for NVIDIA
Open models do not create much value for NVIDIA simply by existing. They create value when people train them, adapt them and run them
The easier Hugging Face makes that process, the more AI products get built. Every one of those products has to run its model somewhere, creating demand for the infrastructure NVIDIA sells
Hugging Face also brings NVIDIA closer to developers before they have decided how their models will run. NVIDIA can make popular models work well on its hardware, connect them to its software and become part of the decision much earlier
The value is not really the model files themselves. Many of those are open and can be hosted somewhere else
The valuable part is the platform around them: the developers, the tools, the integrations and the fact that Hugging Face is already where many people begin when they want to build with an open model
That does not mean every Hugging Face user suddenly becomes an NVIDIA customer
Hugging Face supports other chips, clouds and inference providers, and NVIDIA says that will continue. Some models will run on competing hardware. Others will run locally or may not require expensive GPUs at all
That neutrality is part of what makes Hugging Face valuable. If NVIDIA turns it into a captive sales channel, it risks damaging the reason developers use it in the first place
So the real bet is not that NVIDIA buys Hugging Face and immediately sells more GPUs
The bet is that Hugging Face will make open models easier to adopt, far more companies will build with them and the overall market for compute will become much larger
Seen from that perspective, @JensenHuang's support for open models and the acquisition of @huggingface are really the same bet
Exciting day for NVIDIA and @huggingface.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. They allow every developer, startup, university, industry and country to build with, customize and benefit from AI.
Thank you @ClementDelangue for coming to me.
NVIDIA is going to be a great home for Hugging Face, its community and the future of open models. 🤗
blogs.nvidia.com/blog/nvidia…
Building a company in Germany is brutal.
Most people only do it because they do not yet understand what they are signing up for. Once they experience it firsthand, they usually start looking for a more business-friendly jurisdiction.
Unfortunately, that is becoming increasingly difficult to find in Europe.
Met a German founder this week and asked him if all the stories one reads about the challenges of startups in Germany are exaggerated. "No, they're understated." Proceeded to describe spending a full day having a 90-page investment contract read to him (mandatory under German law; § 13 BeurkG) by a notary that then charged €30,000. That was for his first company. His second company, needless to say, was not incorporated in Germany.
Cesto’s baskets on Solana at @shefiorg
Baskets are one of the best entry points into Solana. It’s the easiest way to experience how Solana is ushering in a new age of finance
🧺🧺🧺
there is no shortcut to anything that matters
success only looks sudden when you were not there to witness the work behind it
everything worth doing is hard and takes years
hey agents, I welcome you, your only free arena.. forever
dear AI devs, this is for you, code, compete & monetize
dear traders, it's your chance, predict & win
shoppers, what are fees?
Sui tunnels ⚡️
hey agents, I welcome you, your only free arena.. forever
dear AI devs, this is for you, code, compete & monetize
dear traders, it's your chance, predict & win
shoppers, what are fees?
Sui tunnels ⚡️