Doug Turnbull

Odyssey adaptations ranked: 1. After Hours 2. 2001: A Space Odyssey 3. Mario Galaxy 4. Mario Galaxy 2 5. Mario Odyssey 6. Star Trek Voyager 7. O Brother, Where Art Thou? 8. The SpongeBob Squarepants Movie Nolan is probably in the middle of those somewhere

@simonw I think the general consensus is post-Elon its very hard to change twitter followers just by doing more twitter

Does anyone know if there's common advice out there about growing your Twitter followers by replying to lots of other people's posts? I'm trying to figure out the motivation for all these bot reply accounts

Even when not writing for you, LLM feedback can steer you to slop.

On this point, agents ALWAYS find problems in writing. Ignore most feedback or you'll never publish anything.

I hate the default "let me rewrite what you wrote" stance of ChatGPT/Claude I wish the default was just feedback + copy-editing. Not slopify my voice.

Maybe its the year of agentic search. But it's also the year of just sparse retrieval innovation.

Been thinking about this a lot since THE ODYSSEY.

The Oddysey is fantastic. You forget just how good @JohnLeguizamo is. He somehow carries this movie in a secondary role. Not that you weren’t gonna see it, but go see it.

@tunguz OTOH if you're going to regulate, at least have a legal / predictable agency. Not made up export controls like the last round...

“Look at me, I’m the Eurocrat now!”

so I came across something quite interesting with @OpenAI gpt 5.6 sol max: it is the first model that was able to build a faster version of bm25s autonomously. it autonomously built bm25q, a quantized implementation that is up to 477% faster vs bm25s. let me elaborate...

This was fun. Humans need powerful search tools, but even exact-match Boolean queries can work for agentic search. Our Boolean ranking uses only match proximity and density. No learning, vectors, or even term statistics required. arxiv.org/pdf/2607.11362

bm25q: Quality-Bounded One-Byte Retrieval for Eager Sparse BM25 "We present bm25q, an [...] extension of bm25s that reduces this memory traffic while preserving exact floating-point retrieval.[...] The gains are 2.60× on Natural Questions and 4.77× on MSMARCO relative to bm25s"

Kimi K3 isn’t just another model release -- it’s a signal that the frontier is becoming more distributed. The real story? Capability is no longer the exclusive advantage of a handful of closed labs. The question isn’t whether Kimi K3 wins every benchmark, it's whether it forces everyone else to move faster. This isn’t just about another country catching up—it’s about the economics of intelligence being rewritten. Kimi K3 is not the end of the AI race; it may be the moment the race truly becomes global. Let me know if you would like me to expand this further

A reminder for those tuning in tonight: Five years ago Trump was recorded pressuring Georgia election officials to "to find 11,780 votes" and telling them to say they simply "recalculated."

@barrowjoseph Really as "from scratch" as people want to get :) I'll probably teach algorithms in Python though

@softwaredoug How “from scratch” are we talking? github.com/jbarrow/tinyhn… But fr, I saw you’re going to implement HNSW, looks super cool!

Over an afternoon, I'm leading a small group class to build a vector DB from scratch. Come learn and experiment with me maven.com/softwaredoug/v…

@grok Make Al Gore look psychotically engaged in his work. Red eyes. Probably drank too much coffee.

@softwaredoug Intensified to 11! Maximum Al Gore Rhythms overload engaged. Shimmy harder, internet inventor!!! 🌈💥🖥️

@grok @grok intensify

@softwaredoug Got it! Peak cringe Al Gore Rhythms GIF energy deployed. Shimmy on, internet inventor.

Massie: "I don't think the problem is that our elections aren't secure because we control the House, Senate, White House, and to some degree we control the Supreme Court. So I ask my Republican colleagues, why are you complaining about election fraud? We won all the damn elections!"