what we learned to keep coding agents fast in Claude Code, Codex, Pi, and more:
– a single timestamp in the system prompt can break the cache
– changing reasoning effort can wipe the whole cache
– telling the agent its output was cut off will get it to read what it missed
coding agents resend the whole conversation on every turn. prefix caching keeps long sessions fast, but only if the harness and the inference engine both do their part.
@zainhas from Together and @parthsareen from @ollama wrote down what each side needs to get right.
together.ai/blog/making-open…
myth retweeted
NEW: An AI researcher at Anthropic runs a business that hosts "consensual non-consent" orgies in the Bay Area. The orgies have been the subject of sexual assault allegations and sparked debate about the AI industry's attitudes toward women, some of whom report feeling pressured to attend the orgies in order to advance their careers.
Paul Crowley, a cybersecurity researcher at Anthropic, is listed as a manager of Red Means No LLC, which says that its parties are "designed to be spaces to … ignore the 'nos' of others." Participants can use a safe word, "red," to end an encounter but give "default consent" to "spanking," "wrestling," and "light pain" by entering the orgy, according to Red Means No's website. More extreme actions, including "light electricity" and "moderate beatings," are "permitted with negotiation," though "body-shaming and race-shaming … are disallowed entirely."
The orgies have been known to attract a disproportionate number of employees at Anthropic and OpenAI, a dynamic critics say creates professional pressure for women in AI to attend. The criticism spilled into public view last month after an independent AI researcher, John Wentworth, said that he "absolutely [does] upgrade" his assessment of AI safety experts "if I find out they go to [Red Means No]," adding that attendance was a "strong signal" of free thought.
"Want to get ahead in your AI career? Go to the rape orgy," Substacker Ivy Asterix quipped. At least one woman claims to have been sexually assaulted at Red Means No, which the group describes as a "safe space" to have "your resistance ignored."🧵
freebeacon.com/america/horri…
myth retweeted
we don't have to live this way. senior engineers have the duty of building guardrails and teaching the next generation of programmers the practices that lead to high quality software.
AIs may write most of the code, but you still decide what ships. we should never lose our taste for quality software.
i made this analogy a while back: everyone can cook at home but yet restaurants still exist. the floor is rising, but so is the ceiling. the best software will always be made by those who care.
myth retweeted
Turned down by many famous publishers
So few people at the top know what they are doing
Even less at the BBC
J.K. Rowling was a single mother living on about £70 a week in benefits when Bloomsbury bought the first Harry Potter book for a £2,500 advance. Several publishers had already turned it down. What changed Bloomsbury's mind was an eight-year-old girl.
Bloomsbury's chairman, Nigel Newton, took the first chapter home and gave it to his daughter Alice.
She came downstairs and told him: "Dad, this is so much better than anything else."
Bloomsbury had only started publishing children's books three years earlier.
Harry Potter and the Philosopher's Stone came out on 26 June 1997. The first print run was 500 copies.
A few months later the US rights went to auction in New York. Scholastic won with a six-figure bid.
"My immediate reaction was to laugh hysterically," she told ITN in 1998. Then: "I felt utter terror."
That first book has since sold around 120 million copies. The seven-book series has sold more than 500 million.
In the same interview, she said it had been turned down by three publishers before Bloomsbury. Any of them could have had it for £2,500.
myth retweeted
The Coldcard hack largely occurred because of the migration away from vetted free and open source trezor code, to libngu so the firmware could be stripped of its FOSS licensing to prevent competition. In their mind, people using their firmware code were parasites despite freely using the trezor code in their product for years.
Hard to imagine a business more hostile to its community and ecosystem.
myth retweeted
Wow, did you even read all that? Of course you didn't.
Fortunately, I put everything into one neat image and a location for you to slide two dollers under my door
ko-fi.com/instoh
Alright, go ahead and clean and report back to me. (9/9)
myth retweeted
Argon, Carbon, and Hydrogen
what does this even mean? Can we please go back to small, medium, and large?
Is argon to Terra as luna is to hydrogen? Or is hydrogen more like Terra? Is argon the luna model? WHAT AM I EVEN DOING HERE?
llama.cpp v1.0.0
github.com/ggml-org/llama.cp…
myth retweeted
We’re not going to win by lobbying, plotting, or manipulating. We’re going to win by building, in the open, with the whole community!
Today, with @ggerganov we’re super excited to announce the release of llama.cpp v1, the first major fully stable version of the most used library to run AI models locally, meaning on your phone, your laptop, and your workstation.
It means you can own your intelligence for free, not send your data to anyone and be 100% in control, wherever you are in the world. It shows our commitment to support every hardware, platform and model type out there and makes AI distributed rather than concentrated in a few Silicon Valley orgs.
Huge thanks to the team, the long-time maintainers and every contributor who helped make this happen. Let’s build! 💛🧡💚
llama.cpp v1.0.0 is now available
huggingface.co/blog/ggml-org…
myth retweeted
I strongly believe DSPy is the best way to build systems with decision models.
✔️ ReAnchor optimizer is cheap, fast, and good
✔️ WeightSpace composes open models and pushes them further
✔️ Signatures let you swap the model you're calling between Jev, alternatives, AND LLMs.
Want to try weight-space programming? Today we're releasing dspy-wsp, a prototype weight-space optimizer that works like any other @DSPyOSS optimizer.
Using dspy-wsp, and about ~$5-10 of Astra, we improved every decision model we tested. cmpnd.ai/blog/weight-space-p…
myth retweeted
Even America’s “haves” are feeling the squeeze.
Prices are rising faster than paychecks: real wages have fallen for six straight months, and 22% of middle-income adults say they can’t fully pay some bills in a typical month, per Pew.
Weekly restaurant-goers have dropped from 53% in 2019 to 39%, per Gallup, and the median first-time homebuyer is now a record 40 years old.
myth retweeted
NEW: I interviewed Matt Lambert repeatedly last year as his bitter condo dispute escalated. He repeatedly criticized building manager James Stuart, whom he killed Thursday.
On the conflict that preceded the Logan Circle shooting: washingtonsun.com/metro/dc/l…
myth retweeted
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo….
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
myth retweeted
this story has everything, including cuckolding
WSJ: "Before the affair that now threatens his job as Nvidia’s chief executive, Jensen Huang was introduced to his future romantic interest, Alice Huang, by a Chinese businessman he’d known for a decade, Kang Wei. Kang also happened to be Alice Huang’s husband, said people who know the couple. For two years, Kang appeared to accept his wife’s romantic relationship with the Nvidia CEO and spoke openly of their closeness, these people said."
wsj.com/tech/ai/a-mystery-bu…































