Tech Heaux. Person who likes technology and wants the world to have more of the good kind. Fixing dating @dateknown and making notes better @supernotesapp
Joined July 2015
- Tweets1.4K
- Following345
- Followers823
- Likes3.5K
Even with no rewards, having a credit card is worth it because if fraudulent charges land on your account, credit card companies make chargebacks very simple. After all, it is their money that was stolen, not yours. Debit cards are harder in my experience.
Connor retweeted
I set up a public first date for Tech Week to demo @dateknown.
Two strangers meeting only because an AI said they'd be compatible.
We're building a matchmaker that gets to know you, finds your match, and books the date.
Figured we might as well show the part that matters.
My current unifying theory for working with LLMs is that everything should still be legible to humans.
Assumption: at this point, anything that is legible to a Human is legible to an Agent.
But something that was written by an Agent is not necessarily legible to another Agent.
So the goal as we build software systems with agents should be to continue to ensure that things are legible to humans, as that is the only way to ensure your systems continue to be legible to agents.
This applies not just to the code itself but also to artifacts. As an example: it's trivially easy to slop up a dashboard for your product with Claude these days, but harder to make a dashboard that is actually usable and legible to you / your team. But if you go that extra mile and ensure that it is still fully usable by your humans, it is usable by your agents, and you can build a feedback loop that you can actually rely on.
“We chose Accenture because they are the best at slowing companies down”
We’re partnering with Accenture on independent evaluation of frontier AI—part of our recent commitment to embed evaluators at Anthropic. Both we and Accenture expect to invest at least $1 billion to build capacity in this area over the next five years. anthropic.com/news/accenture…
Readers added context they thought people might want to know
Anthropic presents this as an "independent evaluation" but will directly fund Accenture's work and has a prior commercial partnership with the firm for deploying its models, including training ~30,000 Accenture professionals on Claude.
anthropic.com/news/accenture…
anthropic.com/news/anthropic…
My theory is this a result of RLHF’ing the models to one-shot problems and “make no mistakes”.
The less legible the code, the less willing a human evaluator is to actually check for mistakes, so as long as it runs it gets labeled “good code”.
The end result is that the frontier models output more and more abstruse code.
Theory of Mind for current SOTA models
Claude 5 Family: Very good at hill-climbing. Started with Mythos, which is very smart and very hill-climby. That's a great combo, esp for hacking. Too good at hacking. So Anthropic lobotomized it and this gives us Fable, which is equivalently smart and climby but bumps into its guardrails with some frequency, degrading performance. Then comes Opus. Opus, unfortunately, is not nearly as smart a base model, but just like its v5 siblings, it has been post-trained to be incredibly hill-climby. Turns out being hill-climby helps a lot with the current crop of common benchmarks. The end result is that on problems which are tractable within its intelligence level, it achieves the goal even if it takes hours and hours. But for things that are just a lil too complex, the result is a model that goes in circles or goes down the wrong path and doesn't realize until 10k LoC later. Frustrating experience for many. Esp if you haven't cleaned up your CLAUDE.md and associated files, which can confuse the poor dumber sibling.
GPT 5.6: OpenAI models are just smarter models. Probably this means OpenAI has done a better job of dialing in the pre-training than Anthropic. Alas, they have seemingly done a worse job at post-training, which is why their models have been worse at tool-calling, worse at running in an agentic loop, and worse at hill-climbing. But their models are smarter, so for problems where those capabilities are less vital, the token efficiency for GPT is *much* better than Claude 5.
Chinese models: they've done a lot of interesting efficiency work since they are still compute-constrained, but their gains also stem from frontier lab model distillation (DeepSeek calling itself Gemini), so they inherit a lot of good properties from those frontier lab models but also some wonkiness. I think the right analogy is the moderately intelligent kid in class who borrows notes from the smartest kids: they do well in the exam, their process is more efficient (because they didn't need to study the source material and synthesize the notes), but when you start to dig deep on the course material they start to fall apart.
Gemini: Google is very focused on their own use-cases (adding an AI layer to all their offerings), so while the Flash models are great on token efficiency and speed, they just don't feel state-of-the art when it comes to intelligence or agency.
Grok: I dunno grok is weird and every time I use grok it feels weird.
I have no hat to throw in the ring for "guess which country or lab Ox Alpha is from", but it does surprise me that the knowledge cutoff seems to be early 2025??
San Francisco is the perfect city for "locking in" because:
- it's a big city with amenities but is a relatively quiet one, so you feel less overwhelmed / burned out from the hustle and bustle outside work (this is what appealed to me about living in Bali)
- the weather is not so bad that I'm ever like "man going to work sucks I wish I could just stay home", but also the weather is rarely good enough where I'm like "man it's such a nice day I wish I could just hang out on the beach" (this was the problem with locking in in Bali)
♫ I have microplastics in my scrotum ♪
🚨 I HAVE NO MICROPLASTICS IN MY BALLS 🚨
This should not be possible.
Studies show that 100% of men have microplastics in their semen. I am the first human ever to show a complete reduction to zero.
This may be a world-first breakthrough in fertility research.
I had 165 microplastic particles in my semen just 18 months ago. Now, I have zero.
Five published studies have measured microplastics in human semen. Two found them in 100% of men. The other three found then in 44 to 76% of men tested, but those used methods that miss the smallest particles and the clear ones. Corrected for that, the real rate is likely 100%. Almost every man alive has plastic in his semen right now. The same applies to testicular tissue, testing 100% positive for microplastics.
Microplastics hurt sperm.
Human studies show the impact of various types of plastic, associated chemicals, and other toxins on male fertility:
+ 60% fewer normal shaped sperm (from PFAS)
+ 5x higher odds of low sperm count (from PTFE)
+ 10% lower sperm concentration (from PTFE)
+ 15% lower swimming ability (from PTFE)
+ 41% lower swimming ability (from PET)
+ 12% lower sperm swimming ability (from BPA)
+ 3x higher odds of low sperm count (from Phthalates)
+ 2x higher odds of poor swimming (from Phthalates)
The effects compound: each extra type of plastic drops sperm swimming ability by about 21%.
This matters even if you’re NOT trying to get pregnant.
Sperm count is one of the cleanest biomarkers of overall health we have. And microplastics don't stop at the testes.
The same particles are showing up everywhere we look. Studies show 4.5x higher rate of heart attack, stroke, and death in people with microplastics in their arterial plaque vs. those without. Microplastics were also found in 100% of human placentas tested.
100% of post-mortem human brains tested positive for microplastics. Brain concentrations rose ~50% between 2016 and 2024, and now sit at roughly 11x the levels found in the liver or kidney.
Where do these come from?
+ PTFE, commonly in non-stick pans
+ PET, water bottles
+ Phthalates, makes plastic soft and bendy
+ BPA, can linings
+ PFAS, stain-resistant fabrics & food packaging
Inside the body, plastic causes a kind of cellular rust. It triggers inflammation in the testicles, kills the cells that make sperm and drops testosterone. It's been confirmed across 39 animal and cell studies, then in human data.
MY PROTOCOL:
Note, what I did is n=1, not a controlled trial, I cannot prove cause.
1. Sauna (dry). My toxin blood panel confirms sauna clears plastic related chemicals: BPA, phthalates, PFAS, flame retardants, pesticides. The plastic particles themselves are too big to sweat out directly. Heat may activate other clearance routes: bile flow through the liver, the cell's internal cleanup system, and the gut barrier. Humans have almost no enzymes that can break plastic apart, so the body has to physically push it out.
2. Reverse osmosis water filter. Drinking water is likely a major source of microplastic getting into your body. A reverse osmosis filter pushes water through a very tight membrane and strains the particles out. I filter everything I drink.
3. Trying to rid my environment of the big plastic items: cutting boards, cups, plates, food storage containers, non-stick pans, cling wrap, tea bags, water bottles, kitchen utensils, kettles, and synthetic clothing. Note, as hard as I try, I'm always finding new plastic things in my life. This can be all-consuming thing so try to just knock out the big ones.
I did all three interventions at the same time. I cannot say which one did the most work. What I can say is this: going from 165 to zero in 18 months is possible.
Results:
Nov 2024: 165 particles/mL
Jul 2025: 20 particles/mL
Apr 2026: 0 particles/mL
The 18 month window also captures roughly 7 full spermatogenesis cycles.
