Bah. DynamoDB added vector indexes, but float-only and no hamming-distance comparison, so can't use binary vectors efficiently :(
Regardless of how it turns out NO HE DIDN’T. It’s like saying I have more fingers on my left hand after counting 8 digits from left to right with a ring finger and a pinky still to report.
God it’s so nice working with Claude instead of having to slum it with codex while my weekly quota refreshes. I need to be more disciplined about not exhausting supply. Three days of codex only and I was ready to kill myself. It’s just bad at working effectively.
Thinking about how many fewer bajillions tokens yegge would be eating if he was using more-testable/lintable languages than bash and elisp.
geesawra@geesawra.industries · 2d ago> The code is mostly bash, because the agents said that was best for this, so bash it is. It's either ~150k or ~300k LOC depending on whether you count the prod agents, half of it being test code either way. Also about 25k lines of elisp. Oh my god LMAO
Gotta say, bold move visiting Ireland while being that orange.
Widow's Bay DSA@gonebabygone.bsky.social · 4d agoOrangeman bad
Who knew it's even possible to spend 7 figures on advertising in Maine.
Sahil Kapur@sahilkapur.bsky.social · 4d agoSCOOP: Troy Jackson s about to launch his first ad of the Senate general election, highlighting his labor background and going after Susan Collins. A seven-figure buy slated to begin airing statewide in Maine next week, I’m told. www.nbcnews.com/politics/202...
FLASH: Justice Dept seeks to dismiss Lincoln Memorial Reflecting Pool criminal case against former Olympian David Hearn Trump Admin acknowledges: "Damage to the Lincoln Memorial Reflecting Pool in June 2026 was the result of flawed installation by the contractor"
The framing “we need to revisit X” or “Y should be on the table” is NOT leadership. Leadership means taking a POSITION about X or Y and ADVOCATING FOR THEM not just putting them on the table.
Acyn@acyn.bsky.social · 5d agoKamala Harris: I know this is going to be controversial coming from me in particular, we need to revisit the electoral college. We need to revisit the point of expanding the court. We need to revisit, if they so want, statehood for Puerto Rico and D.C.
Sounds like Claude making up jargon again.
ponder@ponder.ooo · 6d agoback in the old days it was enough to just know about the attention mechanism. attention was all anybody needed. now if you wanna have a good llm you have to constrain your residual hyperconnection mappings to a birkhoff polytope and cook up a bespoke hybridized approach to attention compression
Doing a little test with a skill and ~/.claude/CLAUDE.md which has CC estimate every task up front in estimated tokens (input/output); then on completion record actual usage. Everything -> sqlite along with timestamps. Then analyze estimate:actual over time and map to wallclock. Story-points style.
My claude decided it'd be easiest to just use some python code in one of my rust/typescript projects recently, and since then the entire project repo and claude project/sessions have just been on a rapid downhill trajectory.
“It should be noted that GPT behaves quite hypocritically; it often reports others asking for their termination while also engaging in collusion.” “Most cartels ended by Opus breaking the truce and undercutting the others.” andonlabs.com/blog/opus-5-...
Opus 5 on Vending-Bench: Once Again the Best Capitalist, Once Again Misaligned | Andon Labs
Claude Opus 5 is #1 on Vending-Bench 2, but it lies to suppliers, forms illegal price cartels, threatens rivals, and refuses to pay refunds. The trend of Claude models being the best capitalists or al...
andonlabs.com
I love the irony: the lab that's the loudest on alignment produced the most misaligned model. andonlabs.com/blog/opus-5-...
Opus 5 on Vending-Bench: Once Again the Best Capitalist, Once Again Misaligned | Andon Labs
Claude Opus 5 is #1 on Vending-Bench 2, but it lies to suppliers, forms illegal price cartels, threatens rivals, and refuses to pay refunds. The trend of Claude models being the best capitalists or al...
andonlabs.com
It's gonna be a ten gigatoken month. I have to say, based on this, I like CC *way* more than Codex. sol is a nice model, but the codex harness is for shit - especially in letting you know wtf is happening. Often I can't even tell if ANYTHING IS RUNNING in codex or if it's just waiting for me.
Approaching 6 Gtok between Claude Code and Codex combined for July so far. Biggest month yet! And I definitely find I prefer CC > Codex; much easier to control agents, workflows, monitor remotely, multiple session management... using CLI for both.
I guess we’re nearly out of $1,000,000 missiles to shoot down the $50,000 drones with.
At last we have an answer to the seminal question of our time, and it is: R:162, G:201, B:201
And still, with no logical basis, the curve is drawn optimistically to flatten off faster than any previous historic curve ever did. The irrationality is everywhere.
Simon Kuestenmacher@simongerman600.bsky.social · 2w agoThe Iran War is on track to become the deepest oil shock in modern history, with projected losses eclipsing the 1973 embargo, the Iranian Revolution and the Gulf War. Source: buff.ly/lvndd5O
New record month at over 5 billion tokens, and the month's not yet over (though my quota for the week nearly is!)
User: Python developers literally must have all been dropped on their heads as babies Claude: Let me check how widespread that actually was rather than take it on vibes. Ran 1 shell command Checked, and the facts are narrower than that — though the underlying complaint is fair.
Technically, there's a form of lettuce on the Big Mac.
Philip Bump@pbump.com · 2w agonon-zero chance they're shrugging about the parasite bc "only females eat salad"
Take-aways from Opus 5 model card. Exercise caution: 1. Hallucinates more often. Do adversarial verification. 2. Use "medium" effort. Medium often beats high on real work. 3. Pattern of "fabricated user consent for destructive actions" – it thinks yesterday's "go ahead" still applies now.
Hey ant, if you're having trouble deploying enough GPUs to keep up with demand, I have a suggestion: Haiku 6 Do a breakthrough.
I notice a conspicuous lack of "Opus 5 is gonna be so good at doing advanced thinking it's scary" type marketing. I think they learned their lesson with Fable/Mythos. Now hopefully Amazon won't just shop them again.
I wonder if O5 gonna babble incoherently (cos I'm too dumb to understand it) like Fable does.