Peter Dedene

I tested Nanbeige4.3-33B on a single DGX Spark: as good with tools as Qwen3.8 Flash Next (BFCL 84.6 / 56.5 vs 82.4 / 54.5) and tied on GPQA. But it's slow: 12 tok/s, even with a 3B drafter And 83 of 200 prompt injections got through (compare to Qwen: just 2). Yes, Qwen stays.

What's broken about Bay Area RL culture right now? Nathan Lambert in Boaz Barak's Harvard class: when an RL run breaks, "you text a friend." When Together AI wrote a paper on logit centering, people said "oh, everybody knew this. It's weird that they wrote a paper on this." 👇

Why open weights still matter for builders: Harvey took a multi-trillion-parameter open model, post-trained it on their own legal evals, and got "a bunch of improvements." Specializing to a niche is "very hard for OpenAI and Anthropic to compete in every vertical."

What's broken about Bay Area RL culture right now? Nathan Lambert in Boaz Barak's Harvard class: when an RL run breaks, "you text a friend." When Together AI wrote a paper on logit centering, people said "oh, everybody knew this. It's weird that they wrote a paper on this." 👇

On distillation: APIs hide chain-of-thought now, and "all the Chinese labs have teams" maintaining methods to "jailbreak the API in order to get these chain-of-thought tokens out." He points to "hundreds of millions of interactions" with the Claude API routed via Singapore.

What's broken about Bay Area RL culture right now? Nathan Lambert in Boaz Barak's Harvard class: when an RL run breaks, "you text a friend." When Together AI wrote a paper on logit centering, people said "oh, everybody knew this. It's weird that they wrote a paper on this." 👇

When the subagent wanders into a folder it has no business being in, touches nothing, and leaves 💀

Why do models learn so much from us when they only train on human data? @geoffreyhinton in Vienna yesterday: a sentence carries "maybe a hundred bits if you're lucky." Identical neural nets averaging gradients share "on the order of a trillion bits." 🤯👇

Multi-agent builders, pay attention to his battery-chicken story. Breeding the top individual hen meant keeping cages "in the dark 90% of the time because the chickens keep killing each other." Breeding the best cage gave cooperative hens. His RL takeaway: "rewarding the groups."

Why do models learn so much from us when they only train on human data? @geoffreyhinton in Vienna yesterday: a sentence carries "maybe a hundred bits if you're lucky." Identical neural nets averaging gradients share "on the order of a trillion bits." 🤯👇

Inside the model, understanding a sentence is like "when a sequence of amino acids folds into a protein." Each word starts as "a thousand weird gradient-learned numbers." Across layers "the words sort of deform each other" until "the interpretations start locking together."

Good breakdown from Theo on what an agent harness is. The fun part is what happens next. Jediah Katz (who built Cursor's and Grok Bot's) says "a fancy harness in this day and age no longer adds coding intelligence." Models absorb those tricks, so you delete that code.

My favorite engineering detail from the talk: they log every deterministic model-behavior hack ("shim") in one place. A bot watches the dashboard, and "whenever a shim stops firing, that model or that agent will put up a PR and say, this shim can be cleaned up."

Good breakdown from Theo on what an agent harness is. The fun part is what happens next. Jediah Katz (who built Cursor's and Grok Bot's) says "a fancy harness in this day and age no longer adds coding intelligence." Models absorb those tricks, so you delete that code.

I tested Recursive Flint (40B, 2.5B active) on one DGX Spark: 93 tok/s, 2x the speed of Qwen3.8 Flash Next, in a third of the memory. Best IFEval I've measured (94.6). But it quits early on long tool chains and thinks Kortrijk is the capital of West Flanders. 🫠 For now, Qwen stays.

Jensen Huang telling little robots to "take a nap" on stage at GTC Berlin is peak 2026. 💀 They're from Pollen Robotics, now part of Hugging Face, and he says they learned their skills in a digital twin on an RTX Spark.

Sam Altman told Matthew Berman he has "not read or even looked at a line of code in probably a year." If he ever writes code by hand again, it'll be "like a revolutionary war reenactment" 💀 Funny line, but somebody still has to read what the agents ship.

He also hates the usual "AI is going to cure cancer" pitch. Dig a little deeper and the price is often "human agency and you in control of the future." What he hopes for instead is that we "avoid this industrial revolution trap" and give people more choice, freedom and time.

Sam Altman told Matthew Berman he has "not read or even looked at a line of code in probably a year." If he ever writes code by hand again, it'll be "like a revolutionary war reenactment" 💀 Funny line, but somebody still has to read what the agents ship.

Asked about OpenAI's moat, he said the old moats keep working. The caveat: if people "really pushed on RSI", the tech could become the moat. He called that "a very dangerous thing to push on" and said the world needs "speed limits or whatever on how quickly we let RSI go."

Sam Altman told Matthew Berman he has "not read or even looked at a line of code in probably a year." If he ever writes code by hand again, it'll be "like a revolutionary war reenactment" 💀 Funny line, but somebody still has to read what the agents ship.

Sir Humphrey's four-stage strategy, php2shell edition: 1. 2024 audit finds the root cause. Nothing will happen. 2. August: reported, RCE confirmed. Something may happen. 3. Ten weeks of patching. 4. It leaks, so the fix ships a day early. Code dating back to 2012, by the way.

Ran the 1.6-bit DeepSeek V4.1 Flash on one DGX Spark over the weekend. It fits (110 of 128 GB) and shrugs off prompt injection: 193 of 200 attacks resisted. But Qwen3.8 Flash Next FP4-Local still wins every row I measured, on 13 GB less memory. Still my pick for a local agent.