
Pinned Tweet
Introducing Regbase: The AI-native Legal Research Platform
Built alongside big law firms, @theregbase cuts 100-hour research projects into work that can be done in just one.
5
10
811
Ethan Shea retweeted
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
Ethan Shea retweeted
Jean-Luc Picard remains the single most compelling vision of progressive politics I am familiar with — secular, rational, compassionate, egalitarian; a believer in progress & abundance but also duty, hard work, and a constant battle for self-improvement and improving the world.
The hardest problem in software engineering:
Choosing a solution roughly as complex as your problem
1
4
146
Some areas are paying for data centers, and others are banning them.
We mapped data center regulations across all 50 states and 1,462 local jurisdictions
My hometown Seattle has a 365-day moratorium!
regbase.com/datacenters/wash…
Before the election, 3 states have paused all new data centers. Many states still have tax incentives to build them.
We built the most a detailed snapshot of data center policy across the country, down to the village level.
Where does your town stand?
1
41
The future of dating is two Claudes flirting
I automated Hinge using Claude Code, a spare Android phone, and a few markdown files.
Getting around 20–25 matches/day, and almost 60% of them are people I would've actually swiped right on myself.
I've been experimenting with this setup for a while now.
Here's a detailed breakdown of how I built it, how it works, and a few things I learned about optimizing my own Hinge profile.
1. The setup is pretty basic
- Took a spare Android phone and installed Hinge.
- Connected it to my MacBook using ADB.
- Connected Claude Code to the phone.
- Gave Claude a prompt with my preferences, deal breakers, non-negotiables, locations, etc.
That's pretty much all the hardware you need.
Also, please don't over-engineer this. Be smart about it.
I know some of you are already thinking about replacing the Android phone with a cloud emulator or some complicated setup. I wouldn't recommend it. You're just increasing the chances of getting flagged or banned.
Keep things simple. There's no point building an entire infrastructure around something that was supposed to save you 45 minutes a day lol.
2. Giving Claude enough context about me
I already have a personality.md file that Claude keeps updating whenever I use it.
It has almost everything about me. Places I've travelled, people I've met, experiences I've had, things that went wrong, things I liked, how I generally think, my preferences, etc.
It's obviously not 100% accurate, but it has a pretty decent understanding of who I am.
I also gave it my entire Hinge profile, including all my photos, prompts, and preferences.
So now, when Claude sees someone's profile, it can actually evaluate whether I'd be interested in that person instead of just blindly swiping.
3. How the actual automation works
I keep travelling between three cities in India, so I've set up three scheduled sessions.
Morning: City 1
Afternoon: City 2
Evening: City 3
Each session runs in batches, with fixed intervals and frequent breaks.
Whenever a profile opens:
Claude reads the basic details first.
If there's an obvious deal breaker, it skips immediately.
Otherwise, it goes through all the photos, prompts, and polls.
Cross-references everything with my preferences.
Decides whether to like or skip.
It's not perfect, but the decisions are surprisingly close to what I would've made myself.
4. The interesting part is how it comments on profiles
I spent some time going through Instagram reels and comment sections about Hinge openers and replies.
Collected around 150 comments that seemed to work well and were also somewhat aligned with my personality.
Put them all into a comments.md file.
Now, whenever Claude decides to like someone's profile, it uses those comments as references.
It doesn't just pick one and paste it.
It reads the entire profile and then writes something specific to that person.
For example, if it's commenting on her first prompt, it might reference something mentioned in her third prompt or one of her photos.
So the comment actually has context instead of being a random pickup line that could be sent to 100 different people.
I think this is one of the reasons the setup has been working reasonably well.
5. Everything gets tracked in a Google Sheet
For every profile, Claude logs:
- Profile details
- Whether it liked or skipped
- Why it made that decision
- What comment it sent
- Whether we eventually matched
The sheet also gets updated whenever a match happens.
I occasionally go through the sheet myself to check how accurate the decisions are.
That's where the ~60% number comes from. Around 60% of the matches are people I would've personally swiped right on.
There's still a lot of room for improvement, but that's not a bad hit rate for something running on autopilot.
6. Something else I tried: optimizing my own Hinge profile
This was actually quite interesting.
Since Claude already has my personality.md, I asked it to analyze my current Hinge profile.
Everything from photos and their sequence to prompts, relationship intentions, and the little descriptions you can add below certain fields.
I also asked it to go through research around dating apps, published studies, and insights shared by people who've worked on products like Hinge.
There are so many small things you generally don't think about.
For example, one claim I came across was that having a black-and-white photo among your regular coloured photos can significantly improve engagement.
The reasoning is that it breaks the visual pattern when someone is scrolling, which might get you a couple of extra seconds of attention.
I haven't independently verified the numbers behind that claim, but it was an interesting thing to test.
Another example is how you describe what you're looking for. Most people just select something like "short-term, open to long-term" and leave it there.
But you can add a small description below it, which gives you another opportunity to communicate what you actually want.
Small things, but they add up.
I've been experimenting with different photo sequences, prompts, and descriptions to see what changes the match rate.
It's basically A/B testing a dating profile lol.
7. Where things stand now
The setup has been running for a while.
Getting around 20–25 matches a day, with roughly 60% matching my own preferences.
Earlier, I'd spend at least 45 minutes a day scrolling through Hinge. With this experiment, the manual effort is almost negligible.
One thing I found interesting is how much difference having proper personal context makes.
Without my personality.md and comments.md, Claude would probably just be another bot sending generic pickup lines.
With those files, it can make decisions and write comments that are much closer to how I'd actually do it.
Again, this was just a fun experiment to see how far I could take Claude Code and phone automation, not something I'm actively using as my regular dating approach. There are obvious questions around authenticity and dating-app rules too.
I've documented the entire setup, including the prompts and files.
If anyone wants the setup file, let me know.
Happy to share it.
Thank me later for saving you a few hours every week.
Now go get back to work :)
Me waiting for a chip startup to build a CPU resistant to side channels by isolating caches, removing ooo execution, prefetching, etc.
Put microcontrollers in the cloud!
Agents can play "Ouija" in the cloud.
This class of timing attacks on shared-host VMs has been known since at least 2015¹. With no full solution.
The world’s cyber infrastructure runs on “yes, that attack is theoretically possible, but who would have the ingenuity & gall to go that far!” Enter, agents.
I'd describe this particular OpenAI find as a 'noisy neighbor' attack. If two agents are on the same machine, one can make enough noise (through the CPU caches!) that the other one can detect it by timing its own computation and cache accesses.
The only "complete" solution would be for each agent to run on its own dedicated hardware, which obviously has significant economic consequences.
For Vercel Sandbox, we've deployed every practical mitigation possible to this class of attack, complementing it with always-on monitoring and red-teaming (similar to where OpenAI arrived).
¹ ieeexplore.ieee.org/document… (and even further back)
64


















