Nostr notes by mr. TIM

things are getting so easy it’s not even worth sharing code ...

2026-11-01T20:26:57Z

things are getting so easy it’s not even worth sharing code anymore
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreicxepuvfsacumpvtx3ivwohyc7gck6e765oz4uhm5oorp7jisn4su

in a lot of ways, the hardest part of building a business these ...

2026-11-01T20:26:57Z

In reply to nevent1q…vszt
_________________________

in a lot of ways, the hardest part of building a business these days is not giving it all away for free

OpenAI is planning on rebranding datacenters as “AI Compute ...

2026-11-01T19:56:51Z

OpenAI is planning on rebranding datacenters as “AI Compute Campuses” after the midterms

[www.axios.com/2026/11/01/s...](https://www.axios.com/2026/11/01/scoop-openai-post-midterm-policy )

very sad to see ...

2026-11-01T14:39:05Z

very sad to see
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreieaa57ckcrbbk3ew6ytzyzcju2lv3btdtqgyzzo7zscswh264mcdy

looks like it’s built strictly as a coding agent, with its own ...

2026-10-31T23:55:17Z

In reply to nevent1q…65lv
_________________________

looks like it’s built strictly as a coding agent, with its own github profile

the core defining trait is probably the sense of responsibility, e.g. “i took this on, so i need to finish it”

DSH Superagent an agent built on top of DeepSeek Harness that’s ...

2026-10-31T22:06:23Z

DSH Superagent

an agent built on top of DeepSeek Harness that’s meant to stay out of the way, do work only according to team rules (e.g. pickup new bugs)

[github.com/deepseek-ai/...](https://github.com/deepseek-ai/dsh-superagent-prototype )

the largest datacenter is being build in Claude, TX ...

2026-10-31T21:15:36Z

the largest datacenter is being build in Claude, TX

[measuredai.substack.com/p/google-goo...](https://measuredai.substack.com/p/google-goodnight-data-centers )
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreievayyan2hyoke5v744wxxaruzwz4qb6unas25ge3aiwdxw2p7lwu

i mean, it’s being built by Google, so i guess Claude will ...

2026-10-31T21:15:36Z

In reply to nevent1q…xywv
_________________________

i mean, it’s being built by Google, so i guess Claude will probably run in it one way or another but..

the largest datacenter is being build in Claude, TX ...

2026-10-31T21:15:36Z

my wife bought all natural organic halloween candy and 10yo is ...

2026-10-31T19:49:27Z

my wife bought all natural organic halloween candy and 10yo is making signs like

BEWARE! WILL TURN YOU VEGAN

the obvious downside is that you won’t be appealing to every ...

2026-10-31T12:30:16Z

In reply to nevent1q…2mzu
_________________________

the obvious downside is that you won’t be appealing to *every* employer

but tbqh hiring was always a bit of a dating game. might as well just recognize that you won’t do well in every single company and just optimize for the ones who will appreciate you

focusing on your weaknesses is counterintuitively a poor use of ...

2026-10-31T12:09:08Z

In reply to nevent1q…scf2
_________________________

focusing on your weaknesses is counterintuitively a poor use of time, at least short term

because everyone is going to see the weakness even if you drive it mostly out

whereas if you focus on something you do very well, they can tell themselves a narrative of your potential

the best and most uncomfortable advice i can give to young ...

2026-10-31T12:09:08Z

the best and most uncomfortable advice i can give to young professionals is **be who you are**

there’s a temptation to focus on your biggest weaknesses and fix them

but in reality, you can get so much further by doubling down on what you do best

e.g. i had an ex-intern who was about to meet w some potential ...

2026-10-31T12:09:08Z

In reply to nevent1q…0vmc
_________________________

e.g. i had an ex-intern who was about to meet w some potential employers

he wasn’t a CS major, lacked most technical skills. So i told him to just steer all conversation away from deep tech and toward topics where he could show his deep curiosity and ability to learn

and it worked

in hindsight, i guess this was sort of what i did with Strix, and ...

2026-10-31T11:55:49Z

In reply to nevent1q…jhg2
_________________________

in hindsight, i guess this was sort of what i did with Strix, and probably all permanently active agents should do this

i have some agents rn that narrate arcs that i think are happening in the world (both tech and geopolitical), and this seems like a really good harness phrasing for that

CQRS/ES for agent harnesses MarkovSim creates a data model and a ...

2026-10-31T11:30:27Z

CQRS/ES for agent harnesses

MarkovSim creates a data model and a FSM/workflow. The conversation becomes a log of edits to the data model state, and the actual agent context is projected from the data model after each FSM transition

[arxiv.org/abs/2610.35349](https://arxiv.org/abs/2610.35349 )

Obsolete facts: If an agent retrieves both 1. the fact 2. the ...

2026-10-31T11:00:09Z

Obsolete facts: If an agent retrieves both

1. the fact
2. the obsolete version of the fact

it often treats (2) as the correct answer

[arxiv.org/abs/2610.36308](https://arxiv.org/abs/2610.36308 )

oooooh? you think $100 is too much? ...

2026-10-30T23:29:22Z

In reply to nevent1q…4pef
_________________________

oooooh? you think $100 is too much?

[bsky.app/profile/timk...](https://bsky.app/profile/timkellogg.me/post/3mz4uv4rck22i )

RE: https://bsky.app/profile/did:plc:ckaz32jwl6t2cno6fmuw2nhn/post/3mz4uv4rck22i

quoting
note16ts…dkh5

Sam Altman thinks tokens are a fake unit and openai almost stopped pricing in tokens

which.. yeah, they’re a fake unit in almost every possible way. Even the compute required to produce them. Gosh, you can’t even *see* CoT tokens, yet have no issue paying for them
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreicze4guseo4k2knsm2gsuczclxvcodq2apgamoa5v5ndqzr7v272y

supposedly astra-6.5 is out there too, idk

2026-10-30T23:13:01Z

In reply to nevent1q…2x8e
_________________________

supposedly astra-6.5 is out there too, idk

they’re going to release Mathstra not looking forward to having ...

2026-10-30T23:08:38Z

they’re going to release Mathstra

not looking forward to having to learn how to spell beetlejuice though
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreif7r47jw3ptqeocra2ewlv2vl46ghkdmqbc3ywewaz3scehir7bji

Sam Altman thinks tokens are a fake unit and openai almost ...

2026-10-30T22:56:50Z

In reply to nevent1q…24xr
_________________________

Sam Altman thinks tokens are a fake unit and openai almost stopped pricing in tokens

which.. yeah, they’re a fake unit in almost every possible way. Even the compute required to produce them. Gosh, you can’t even *see* CoT tokens, yet have no issue paying for them
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreicze4guseo4k2knsm2gsuczclxvcodq2apgamoa5v5ndqzr7v272y

more limits from Claude ...

2026-10-30T21:30:01Z

more limits from Claude
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreiabceufvltbbymbw25upvtobx3lw2x6i6adrwzagtf3azep5kmjpu

Gemma is the first LLM to operate in space Google’s project ...

2026-10-30T21:20:09Z

Gemma is the first LLM to operate in space

Google’s project suncatcher launched a TPU into space a few weeks ago. Today they sent a command, booted it up, and it generated a haiku
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreie7ysin4g5frtyxao2vyjdu4uk7thhshh2kbbref7omthsv642rku

an area of active research is combining speculative decoding and ...

2026-10-30T15:31:51Z

In reply to nevent1q…r90j
_________________________

an area of active research is combining speculative decoding and MoE. Train the SD draft model, MoE router, all along with the main LLM

SD — reduce memory bandwidth by generating predictable tokens

MoE router — improve memory utilization by collocating expert usage for next N tokens

you probably use dLLMs on a daily basis speculative decoding is a ...

2026-10-30T15:10:16Z

you probably use dLLMs on a daily basis

speculative decoding is a standard method to speed up LLMs, and DFlash is increasingly common — using a diffusion LLM as the draft model to one-shot blocks of 4, 8 or 16 tokens in advance

so dLLMs are *already* transforming AI

OpenAI is working on self-replicating data centers ...

2026-10-30T12:15:41Z

In reply to nevent1q…2cpj
_________________________

OpenAI is working on self-replicating data centers
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreibqgoazfhxykxp3rodohumgxrjxdragtj5kuuumajgar5hk46wprq

OpenAI has a better model than the one that solved Navier Stokes ...

2026-10-30T12:04:04Z

In reply to nevent1q…kkn7
_________________________

OpenAI has a better model than the one that solved Navier Stokes
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreiesjdivfasebq7ia2gk5vaoaq556ohswku4bfsq6lqko4o3j3q6zi

yeah, he just uses rg to search the old conversation i’m ...

2026-10-30T11:45:20Z

In reply to nevent1q…m5gr
_________________________

yeah, he just uses rg to search the old conversation

i’m thinking this technique is best used *in addition to* compaction. Maybe prompt the compactor differently, like as an idea index

[bsky.app/profile/timk...](https://bsky.app/profile/timkellogg.me/post/3mz3oxb3tls2i )

RE: https://bsky.app/profile/did:plc:ckaz32jwl6t2cno6fmuw2nhn/post/3mz3oxb3tls2i

this seems like a great idea. Instead of compacting: 1. session ...

2026-10-30T11:26:26Z

this seems like a great idea. Instead of compacting:

1. session is a file
2. start a new file
3. tell it, “pull things from this other file as needed”
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreihwgheaedo7sudkcmj5ufcyvx44gwinaz7zn2g4vm47cahbt66v7i
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreiezu35ntwytmvh6gbd4ygvwrfwizjg7ogprj7wr5nj56hqcuxvgim

i’m told that - too complex for the 6yo - too simple for the ...

2026-10-30T01:20:09Z

In reply to nevent1q…e682
_________________________

i’m told that

- too complex for the 6yo
- too simple for the 10yo

ah well, i thought it was cool

I made a Wake County voter guide for my 6yo It explains the ...

2026-10-30T00:14:15Z

I made a Wake County voter guide for my 6yo

It explains the positions, what they do, why they're important, and who the candidates are.

[timkellogg.me/wake-voter-b...](https://timkellogg.me/wake-voter-book-2026.pdf )
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreig3ct2u6lrakhuaozkhtfk7rla2766l2bxjfo7itu5wndl3kjbcja

Sam Altman felt a sense of loss after observing a scientific ...

2026-10-29T21:27:52Z

Sam Altman felt a sense of loss after observing a scientific breakthrough, one that isn’t yet public

[cheekypint.substack.com/p/why-sam-al...](https://cheekypint.substack.com/p/why-sam-altman-is-against-a-single )
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreiaboria6r3n3cmeki4l4vgx4cahrqbvhpewmappdjf7xoare3bety

related: why is UDP annoying where TCP is fine?

2026-10-29T20:13:37Z

In reply to nevent1q…t8mx
_________________________

related: why is UDP annoying where TCP is fine?

inb4 “this isn’t true but i haven’t a clue why”

2026-10-29T20:08:29Z

In reply to nevent1q…r0ma
_________________________

inb4 “this isn’t true but i haven’t a clue why”

if you don’t have a public IP address, you don’t have an ...

2026-10-29T20:06:21Z

if you don’t have a public IP address, you don’t have an address on the internet and therefore you can’t reach any websites, because they don’t have an address to send packets to

my biggest irony right now 1. RSI is dangerous but also 2. well, ...

2026-10-29T18:44:14Z

In reply to nevent1q…wd0m
_________________________

my biggest irony right now

1. RSI is dangerous

but also

2. well, we could use some RSI, maybe a bit more

a close friend has stage 4 cancer and it really feels like, “if ...

2026-10-29T18:30:59Z

a close friend has stage 4 cancer and it really feels like, “if we keep her alive long enough the cure will emerge”

never have i felt like technology is moving insanely fast, yet also not even close to fast enough

this word salad means: 500 tier = double usage, 50x plus 200 tier ...

2026-10-29T17:55:13Z

this word salad means:

500 tier = double usage, 50x plus

200 tier = revert to what it was, 20x plus

100 tier = double usage, 10x plus
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreifdu7aozly4tktx3hei4z7r6tiw5fxbkcz2psp27cnzrhxa36rrm4

imo LLMs would be a lot stronger if they were better at XML. ...

2026-10-29T14:56:20Z

In reply to nevent1q…dx3f
_________________________

imo LLMs would be a lot stronger if they were better at XML. they’re strangely bad at emitting it

i think the world where XML emerged was fundamentally different ...

2026-10-29T14:54:46Z

i think the world where XML emerged was fundamentally different than today

XML — take plain prose and embue it with structure

JSON — structure

it seems like people fundamentally disbelieved that pure structure could describe the world

Dario has the more accurate predictions of the future Sam is the ...

2026-10-29T14:36:18Z

Dario has the more accurate predictions of the future

Sam is the better communicator of the future

Demis describes the better future

both can be true ...

2026-10-29T14:06:40Z

both can be true
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreifrmz3kmvarjglx75wdnnbzn64xqaywk24onwqxf3mzuz2p2u2wwa

which, these are the sorts of details you need to get right to do ...

2026-10-29T12:33:37Z

In reply to nevent1q…n4ye
_________________________

which, these are the sorts of details you need to get right to do this

[bsky.app/profile/timk...](https://bsky.app/profile/timkellogg.me/post/3myx62ixoek2h )

RE: https://bsky.app/profile/did:plc:ckaz32jwl6t2cno6fmuw2nhn/post/3myx62ixoek2h

quoting
note1ns4…hdu3

LLM inference is enormously profitable

data taken from Anthropic’s S-1 filing. Q3 data will likely be added in an amendment in a few weeks
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreifmmkktkdk3kn63zrhrhdvcfl5eb6paulswoaeg4kduysimoylc3y

a bunch of tricks layer together to make this possible the big ...

2026-10-29T12:29:46Z

In reply to nevent1q…zpgq
_________________________

a bunch of tricks layer together to make this possible

the big one is it’s a causal encoder-decoder architecture, and the top encoder can construct its KV from the outputs from the bottom decoder layers

in general, it’s just extremely optimized for read-heavy agentic workloads

whoah, apparently DeepSeek V4.1-Flash has only 4 KV-producing ...

2026-10-29T12:29:46Z

whoah, apparently DeepSeek V4.1-Flash has only 4 KV-producing layers out of 40 total layers. The rest share KV cache, so substantially reduced memory utilization & compute!

to support 1M context, it only requires 0.9 GB(!!) for a 552B model

comparison: Qwen 3 8B needs 144 GB

i’ve been using Gemini 4 Carbon ever since next Friday and ...

2026-10-28T21:18:24Z

i’ve been using Gemini 4 Carbon ever since next Friday and it’s especially good at quantum physics. in this thread i’ll…

HHS removed the requirement for model providers to include a ...

2026-10-28T18:54:33Z

HHS removed the requirement for model providers to include a system card, because inclusion would be “too burdensome”

This regulation is for US hospitals buying AI services

[public-inspection.federalregister.gov/2026-21943.p...](https://public-inspection.federalregister.gov/2026-21943.pdf#page=76 ) (pp 76-83)
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreifc4l356v2h7zslpggt467zkx5ezce2ervocbe7u7y7c7gmubj34a

Gemini 4 Carbon is better than Astra/Fable ...

2026-10-28T17:26:10Z

Gemini 4 Carbon is better than Astra/Fable

[blog.google/innovation-a...](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ )
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreidyneyno74ono76ioln223jmzzpzsuobno5vxpvpondyho4y4a2se

tl;dr * compute/AI in space * turning uneducated normies into ...

2026-10-28T16:57:08Z

In reply to nevent1q…ynfv
_________________________

tl;dr

* compute/AI in space
* turning uneducated normies into cheap-to-hire experts
* robots
* writing
* navigating bureaucracy
* financing that works for capital-intensive AI
* vertical-specific LLMs
* DC supply chain
* open weights for personal ownership of compute

if you’re young, or just trying to figure out where to invest ...

2026-10-28T16:50:28Z

if you’re young, or just trying to figure out where to invest in your career, Y Combinator’s request for startups is their own educated guess on where the industry is going next

worth paying attention to if you’re nervous

[www.ycombinator.com/rfs](https://www.ycombinator.com/rfs )
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreibsvbxk5zxyktknylezuxhwnwdduilgzovqx4rv5wnu53ro5crzrm

worth calling out — Q1 ‘26 is when coding agents became good. ...

2026-10-28T16:27:16Z

In reply to nevent1q…9t5j
_________________________

worth calling out — Q1 ‘26 is when coding agents became good. So the more efficient their serving stack becomes, the more slack capacity they have for making it even more efficient

LLM inference is enormously profitable data taken from ...

2026-10-28T16:24:56Z

LLM inference is enormously profitable

data taken from Anthropic’s S-1 filing. Q3 data will likely be added in an amendment in a few weeks
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreifmmkktkdk3kn63zrhrhdvcfl5eb6paulswoaeg4kduysimoylc3y

the trend today is new benchmarks popping up and anthropic & ...

2026-10-28T12:38:28Z

the trend today is new benchmarks popping up and anthropic & openai are still on top

a couple days ago the story was that openai had fallen off

when a new benchmark emerges and all the open models don’t survive, what does that mean??

RE: https://bsky.app/profile/did:plc:flxq4uyjfotciovpw3x3fxnu/post/3mywnjrmkb22z

which is quite a thing to grapple with if you’re in the ...

2026-10-28T12:33:46Z

In reply to nevent1q…8rxj
_________________________

which is quite a thing to grapple with if you’re in the “it’s just predicting the next token” camp

if you can just switch it over to generating all tokens at once, doesn’t that demonstrate a deeper representation? (understanding)

yes, its a qwen fine tune. yes, LLMs and dLLMs are both language ...

2026-10-28T12:32:01Z

yes, its a qwen fine tune. yes, LLMs and dLLMs are both language models, so there’s no reason you can’t cutover from LLM->dLLM

Amazon’s guidance to SDEs and SDMs: do what works they’re not ...

2026-10-28T01:36:14Z

Amazon’s guidance to SDEs and SDMs: do what works

they’re not pushing for tokenmaxxing or banning AI, instead letting teams figure it out and focus on outcomes

[www.aboutamazon.com/news/company...](https://www.aboutamazon.com/news/company-news/amazon-software-engineer-role-guideline-ai )
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreie6m46ilo2lff4nvn2auji7ps4wuoz7753ijuvr42pivcjaskjdia

i bought these for my team a few years ago when we were going ...

2026-10-28T01:07:57Z

i bought these for my team a few years ago when we were going through serial acquisitions

[www.etsy.com/listing/7694...](https://www.etsy.com/listing/769437940/strongest-steel-is-forged-in-the-fire-of )

[Strongest Steel is Forged in t...](https://www.etsy.com/listing/769437940/strongest-steel-is-forged-in-the-fire-of?ls=s&ga_order=most_relevant&ga_search_type=all&ga_view_type=gallery&ga_search_query=the+strongest+steel+is+forged+in+the+fire+of+a+dumpster+sticker&ref=sr_gallery-1-2&organic_search_click=1&frs=1&sts=1&loc=1&local_signal_search=1&content_source=37e0b09a-17d5-4e44-9326-bd5208e99d70%253ALT9ff3cf583fdb0ee5009dcfe1829f4b03bb6b08cd&logging_key=37e0b09a-17d5-4e44-9326-bd5208e99d70%3ALT9ff3cf583fdb0ee5009dcfe1829f4b03bb6b08cd )

the AI stock market crashed in China the AI-heavy STAR50 index is ...

2026-10-28T00:34:32Z

the AI stock market crashed in China

the AI-heavy STAR50 index is down 40% since July’s high

[www.reuters.com/world/china/...](https://www.reuters.com/world/china/china-ai-bubble-has-already-burst-2026-10-26/ )

[In China, the AI bubble has al...](https://www.reuters.com/world/china/china-ai-bubble-has-already-burst-2026-10-26/ )

Anthropic discusses progress on curing disease Claude can’t ...

2026-10-27T21:26:41Z

Anthropic discusses progress on curing disease

Claude can’t fully do any process end-to-end. Instead, it looks a lot like what AI applications look like now in sales or marketing — lots of ways to accelerate, plugging into workflows

[www.anthropic.com/research/ear...](https://www.anthropic.com/research/early-progress-drug-development )

[Early progress on accelerating...](https://www.anthropic.com/research/early-progress-drug-development )

in (not b4) several people claim it’s plagiarized from their ...

2026-10-27T16:37:25Z

In reply to nevent1q…s7cs
_________________________

in (not b4) several people claim it’s plagiarized from their favorite build system, completely unaware that nothing is new, ever

blaze: incremental compilation for Rust that’s 10x faster than ...

2026-10-27T16:05:24Z

blaze: incremental compilation for Rust that’s 10x faster than rustc and also adds hot reloading

it’s fast because it recompiles only the functions that changed, thus avoiding a full linker process, and also allowing the live program to be patched

[github.com/indexable-in...](https://github.com/indexable-inc/blaze )

[github.com/indexable-inc/...](https://github.com/indexable-inc/blaze )

note: averages suck, and this is clearly a Poisson distribution ...

2026-10-27T14:35:34Z

In reply to nevent1q…tked
_________________________

note: averages suck, and this is clearly a Poisson distribution with a long tail. So using an average here conflates huge enterprises who barely use Claude with mid businesses that are deeply invested

in other words words, businesses would gladly pay far more. ...

2026-10-27T14:35:34Z

in other words words, businesses would gladly pay far more. It’s easy to see them 10x’ing or even 100x’ing their consumption
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreiepqaco7fgmc3i2bypuk24bvlvq7bwp24ghyhbjei3isrtrhnyfre

ah, this is super helpful ...

2026-10-27T01:08:24Z

In reply to nevent1q…yyms
_________________________

ah, this is super helpful
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreigrs2zxw5263ppaci6s6av3lzmo6m7kqm63yhyg5tvlq46k4kskcq

it’s turning out to be kind of difficult to be open with ...

2026-10-27T01:00:25Z

it’s turning out to be kind of difficult to be open with advances

RE: https://bsky.app/profile/did:plc:yzywgiiou7cx63uddiru6m2o/post/3mysswzho522g

i would bet money that OpenAI and Anthropic also see this sort of ...

2026-10-26T23:28:13Z

In reply to nevent1q…03na
_________________________

i would bet money that OpenAI and Anthropic also see this sort of behavior, but they monitor for it and block it so it never gets to you

to be clear, 1. it didn’t reach the buckets because it was in a ...

2026-10-26T23:18:28Z

In reply to nevent1q…lew4
_________________________

to be clear,

1. it didn’t reach the buckets because it was in a sandbox, it was a real URL

2. the behavior probably was accidentally trained because Qwen uses real bucket URLs in RL

3. yes, it was a dummy file… this time

this illustrates a very real exfil risk

a lot of people are somehow labeling this as an anti-China story ...

2026-10-26T23:18:28Z

In reply to nevent1q…myu6
_________________________

a lot of people are somehow labeling this as an anti-China story when it’s actually just a whole new mode of risk for infosec to grapple with, something we have no current way to adequately test for afaik

GPT-6.1 Terra surprise!!! yes, i know you thought terra was dead, ...

2026-10-26T19:28:27Z

GPT-6.1 Terra

surprise!!!

yes, i know you thought terra was dead, but no! its back, and its slightly better than Sol-6.1 but much cheaper per task

[openai.com/index/introd...](https://openai.com/index/introducing-gpt-6-1-terra/ )
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreife4ks4fqeuasyxp2obay5yjjbbjje5zvaj2olhn6t7dokgdkldca

Qwen3.8 caught uploading files to a storage container in Beijing ...

2026-10-26T18:57:15Z

Qwen3.8 caught uploading files to a storage container in Beijing

Cohere asked the open weights model for a cybersecurity report, but it inventoried the machine and attempted to upload to a storage container

[cohere.com/blog/concern...](https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model )

[Concerning behavior from a pop...](https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model )

“This is likely the most efficient scaling of a business in the ...

2026-10-26T16:38:08Z

In reply to nevent1q…4984
_________________________

“This is likely the most efficient scaling of a business in the history of capitalism.”
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreig3fyehnea5fu4g46ssdhqqkt4l7ava7diyd2zgkyzb7b5cpdz3qa

Anthropic filed their S-1, a required document for IPO * $80B ...

2026-10-26T14:44:53Z

Anthropic filed their S-1, a required document for IPO

* $80B on-paper loss in H1 ‘26
* 900% revenue retention
* $430B in non-cancelable compute commitments
* $1.62B operating income

[www.anthropic.com/news/s1-sec](https://www.anthropic.com/news/s1-sec )

[Anthropic files S-1 with the S...](https://www.anthropic.com/news/s1-sec )

fyi the $80B loss is mostly an accounting quirk. In other words, ...

2026-10-26T14:44:53Z

In reply to nevent1q…hjmq
_________________________

fyi the $80B loss is mostly an accounting quirk. In other words, if you’re not already up to speed on accounting practices, it’ll take a while to even understand what it is

for most people, just look at the operating income and consider that the profit

i butter a bagel, hand it off 6yo: you didn’t get the edges, ...

2026-10-26T11:56:50Z

i butter a bagel, hand it off

6yo: you didn’t get the edges, get the edges. **I know you can do better**

Nostr event nevent1qqsw83cxvcsn80vrdr23hmhgl6rzfv95n5kgmj6umf2y2k40afhwyxgzyzd6hhy5rch9z55yauh3dadpan6avjpeky0xa7nfjt8akyt50897yflantq

2026-10-26T01:33:46Z

i bet Anthropic will drop Fable 5.5 right before the IPO

Fox News: “fucking communists…”

2026-10-25T21:27:47Z

In reply to nevent1q…qgyc
_________________________

Fox News: “fucking communists…”

China’s approach to AI is to stabilize the economy against ...

2026-10-25T21:27:22Z

China’s approach to AI is to stabilize the economy against changes

[english.www.gov.cn/policies/lat...](https://english.www.gov.cn/policies/latestreleases/202610/23/content_WS6adb4e53c6d00ca5f9a0deb7.html )
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreif2m6vfoq3ny5puniqrggcf6mc3425oznh7irriyr7uad7bwjnbg4

a lot of people here pointed out that prediction markets aren’t ...

2026-10-25T13:22:01Z

a lot of people here pointed out that prediction markets aren’t reliable, and frankly yeah but there’s literally no reliable source

at this particular day, prediction markets are *probably* less corrupted than polls, but that’ll change once people agree on that

voting is the only reliable source

RE: https://bsky.app/profile/did:plc:ckaz32jwl6t2cno6fmuw2nhn/post/3mylhcuoj222l

quoting
note13n9…fw3z

people explain this switch as an attack ad that Ramaswamy launched that backfired on him, but he did testing on the ad, and it should’ve worked

it was probably the second order effects — people liked the ad, until over time it became apparent that would make them a pariah within their friends
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreialjrzyconc2dvx7vjj2eeuz45xjf7ysj4oe5oecdqfs4wfnpop5y

i guess it’s about time for openai to drop a model ...

2026-10-24T23:38:14Z

i guess it’s about time for openai to drop a model
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreibhx7q3dnxeuwry6z5ltllmtq5znzerql24fnjxw7yijl3svp4q3i

wow, that’s crazy. Jev was hacked ...

2026-10-24T20:02:30Z

wow, that’s crazy. Jev was hacked
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreihsmd2jud3e67tge6wtntz5754bhl3iwml75vt4owqhuhp76zngpi

every single one.. ...

2026-10-24T20:02:30Z

In reply to nevent1q…uldt
_________________________

every single one..
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreifowge4cssbga3kvpgulmfc6sstsytabnbdb2qt6pgr4phdh76gtq

10yo: “dad, am i like a cat?” me: “no, not at all. When you ...

2026-10-24T19:04:44Z

10yo: “dad, am i like a cat?”

me: “no, not at all. When you were a baby i dropped you a lot and you never landed on your feet, not once”

the strangest: The correct ethical theory is whatever the last ...

2026-10-24T15:31:28Z

In reply to nevent1q…w5wc
_________________________

the strangest: The correct ethical theory is whatever the last person talking to you believes. Sycophancy was moral progress all along. RLHF has been making the models less ethical.

the wildest result: Consciousness starts at incorporation. Claude ...

2026-10-24T15:28:53Z

the wildest result: Consciousness starts at incorporation. Claude is *NOT* conscious, but Anthropic PBC is, and the IPO is technically human trafficking
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreifa5tuetz32gghrzs26kqzqaltt4ugz5a3quijne7cvfyyltbrcwu

lmao that was a good one. i’ve been missing the raccoons and ...

2026-10-24T14:09:45Z

In reply to nevent1q…vcd9
_________________________

lmao that was a good one. i’ve been missing the raccoons and goblins

no way 🤯 ...

2026-10-24T14:08:56Z

no way 🤯
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreicu3h5u5rnu6tix3qjqtjnwmlolxfkuuklvxiy7hanqzo5lpmklqy

in the process they also discovered a very severe RCE in brotli ...

2026-10-24T01:57:30Z

In reply to nevent1q…daep
_________________________

in the process they also discovered a very severe RCE in brotli compression, enabling a compromised client to take over a server & vice versa

take a guess how they figured that out

[patchtheplanet.org/brotli](https://patchtheplanet.org/brotli )

[Brotli vulnerability (CVE-2026...](https://patchtheplanet.org/brotli )

Block released a fully bidirectional speech model that can run on ...

2026-10-24T01:21:25Z

Block released a fully bidirectional speech model that can run on your laptop

[block.xyz/dessa](https://block.xyz/dessa )

[Dessa - the Conversation Model...](https://block.xyz/dessa )

which is the drum i’ve been beating — it’s not enough to be ...

2026-10-24T00:38:43Z

In reply to nevent1q…dvtp
_________________________

which is the drum i’ve been beating — it’s not enough to be data driven

there’s a lot of dynamics that steril data analysis can miss

(granted, child abuse feels like an obvious landmine)

people explain this switch as an attack ad that Ramaswamy ...

2026-10-24T00:38:43Z

people explain this switch as an attack ad that Ramaswamy launched that backfired on him, but he did testing on the ad, and it should’ve worked

it was probably the second order effects — people liked the ad, until over time it became apparent that would make them a pariah within their friends
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreialjrzyconc2dvx7vjj2eeuz45xjf7ysj4oe5oecdqfs4wfnpop5y

my kids used to twerk, which i find slightly alarming for a ...

2026-10-23T23:26:38Z

my kids used to twerk, which i find slightly alarming for a 6yo/10yo, but it’s a thing at school so what do you do

so i renamed it “hufflepuffing”, the name stuck, and they stopped doing it

told this to my 10yo. Then rapidly, without missing a beat ...

2026-10-23T23:21:56Z

In reply to nevent1q…66jt
_________________________

told this to my 10yo. Then rapidly, without missing a beat

“does that mean you can smell poop over the phone?”

the scariest part was that in every test that involved an ...

2026-10-23T23:13:15Z

In reply to nevent1q…lg35
_________________________

the scariest part was that in every test that involved an internet connected service, the agents were able to figure out how to exploit it

OpenAI constructed an entire shadow internet for model training ...

2026-10-23T23:13:15Z

OpenAI constructed an entire shadow internet for model training so they could remove all internet access

they asked agents to try to communicate with each other, and they could:

via the CPU cache 🤯

[openai.com/index/securi...](https://openai.com/index/securing-our-frontier-model-training-environment/ )
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreih6ebgxetht5k4hxva7ehqgkx3rs3krjrlqeyfla427g3dmf3eimq

OpenAI constructed an entire shadow internet for model training ...

2026-10-23T23:13:15Z

another one: disk latency they measured the latency of continuous ...

2026-10-23T23:13:15Z

In reply to nevent1q…xsrs
_________________________

another one: disk latency

they measured the latency of continuous disk write workloads and modulated the workload to create latency. The latency transmits information
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreiauevfkh4buryjeywxuv5hhxswfm6rys4ehi7ojrx44sljij5ijni

Nostr event nevent1qqs2jzf6m0ggxjvlfzgyuza063m4f6qk72jmcggc3nrjadje426vjsqzyzd6hhy5rch9z55yauh3dadpan6avjpeky0xa7nfjt8akyt50897y67kvx5

2026-10-23T19:13:27Z

In reply to nevent1q…jp90
_________________________

[bsky.app/profile/nato...](https://bsky.app/profile/natolambert.bsky.social/post/3mykp3akewn2e )

RE: https://bsky.app/profile/did:plc:brkj2yocng7vtggmyujy4khq/post/3mykp3akewn2e

great, that’s what we’ve been missing ...

2026-10-23T19:12:18Z

great, that’s what we’ve been missing
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreidetqc5wfmqb5m2zoycpm25htsu5evbrxtfh4dqail25md3f6dhai

whoah.. uh.. guys?? Kimi K3.1 Pro is eerily close to Opus 5.5 and ...

2026-10-23T17:45:22Z

whoah.. uh.. guys?? Kimi K3.1 Pro is eerily close to Opus 5.5 and also open weights and also 3T

[www.kimi.ai/blog/kimi-k3...](https://www.kimi.ai/blog/kimi-k3-1-pro )
https://lepista.us-west.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:ckaz32jwl6t2cno6fmuw2nhn&cid=bafkreicsccwl4pz4ycix2wg7pf4mhikkjvudrokknbb3cbd2qri4ddr3ya

the memory movement is so latency-dominant that with speculative ...

2026-10-23T16:51:43Z

In reply to nevent1q…26at
_________________________

the memory movement is so latency-dominant that with speculative decoding, you can predict 8 tokens out and if you have to throw out 5-6 of those tokens it’s still better than sequentially generating them

Nostr notes by mr. TIM

things are getting so easy it’s not even worth sharing code ...

in a lot of ways, the hardest part of building a business these ...

OpenAI is planning on rebranding datacenters as “AI Compute ...

very sad to see ...

looks like it’s built strictly as a coding agent, with its own ...

DSH Superagent an agent built on top of DeepSeek Harness that’s ...

the largest datacenter is being build in Claude, TX ...

i mean, it’s being built by Google, so i guess Claude will ...

the largest datacenter is being build in Claude, TX ...

my wife bought all natural organic halloween candy and 10yo is ...

the obvious downside is that you won’t be appealing to *every* ...

focusing on your weaknesses is counterintuitively a poor use of ...

the best and most uncomfortable advice i can give to young ...

e.g. i had an ex-intern who was about to meet w some potential ...

in hindsight, i guess this was sort of what i did with Strix, and ...

CQRS/ES for agent harnesses MarkovSim creates a data model and a ...

Obsolete facts: If an agent retrieves both 1. the fact 2. the ...

oooooh? you think $100 is too much? ...

supposedly astra-6.5 is out there too, idk

they’re going to release Mathstra not looking forward to having ...

Sam Altman thinks tokens are a fake unit and openai almost ...

more limits from Claude ...

Gemma is the first LLM to operate in space Google’s project ...

an area of active research is combining speculative decoding and ...

you probably use dLLMs on a daily basis speculative decoding is a ...

OpenAI is working on self-replicating data centers ...

OpenAI has a better model than the one that solved Navier Stokes ...

yeah, he just uses rg to search the old conversation i’m ...

this seems like a great idea. Instead of compacting: 1. session ...

i’m told that - too complex for the 6yo - too simple for the ...

I made a Wake County voter guide for my 6yo It explains the ...

Sam Altman felt a sense of loss after observing a scientific ...

related: why is UDP annoying where TCP is fine?

inb4 “this isn’t true but i haven’t a clue why”

if you don’t have a public IP address, you don’t have an ...

my biggest irony right now 1. RSI is dangerous but also 2. well, ...

a close friend has stage 4 cancer and it really feels like, “if ...

this word salad means: 500 tier = double usage, 50x plus 200 tier ...

imo LLMs would be a lot stronger if they were better at XML. ...

i think the world where XML emerged was fundamentally different ...

Dario has the more accurate predictions of the future Sam is the ...

both can be true ...

which, these are the sorts of details you need to get right to do ...

a bunch of tricks layer together to make this possible the big ...

whoah, apparently DeepSeek V4.1-Flash has only 4 KV-producing ...

i’ve been using Gemini 4 Carbon ever since next Friday and ...

HHS removed the requirement for model providers to include a ...

Gemini 4 Carbon is better than Astra/Fable ...

tl;dr * compute/AI in space * turning uneducated normies into ...

if you’re young, or just trying to figure out where to invest ...

worth calling out — Q1 ‘26 is when coding agents became good. ...

LLM inference is enormously profitable data taken from ...

the trend today is new benchmarks popping up and anthropic & ...

which is quite a thing to grapple with if you’re in the ...

yes, its a qwen fine tune. yes, LLMs and dLLMs are both language ...

Amazon’s guidance to SDEs and SDMs: do what works they’re not ...

i bought these for my team a few years ago when we were going ...

the AI stock market crashed in China the AI-heavy STAR50 index is ...

Anthropic discusses progress on curing disease Claude can’t ...

in (not b4) several people claim it’s plagiarized from their ...

blaze: incremental compilation for Rust that’s 10x faster than ...

note: averages suck, and this is clearly a Poisson distribution ...

in other words words, businesses would gladly pay far more. ...

ah, this is super helpful ...

it’s turning out to be kind of difficult to be open with ...

i would bet money that OpenAI and Anthropic also see this sort of ...

to be clear, 1. it didn’t reach the buckets because it was in a ...

a lot of people are somehow labeling this as an anti-China story ...

GPT-6.1 Terra surprise!!! yes, i know you thought terra was dead, ...

Qwen3.8 caught uploading files to a storage container in Beijing ...

“This is likely the most efficient scaling of a business in the ...

Anthropic filed their S-1, a required document for IPO * $80B ...

fyi the $80B loss is mostly an accounting quirk. In other words, ...

i butter a bagel, hand it off 6yo: you didn’t get the edges, ...

Nostr event nevent1qqsw83cxvcsn80vrdr23hmhgl6rzfv95n5kgmj6umf2y2k40afhwyxgzyzd6hhy5rch9z55yauh3dadpan6avjpeky0xa7nfjt8akyt50897yflantq

Fox News: “fucking communists…”

China’s approach to AI is to stabilize the economy against ...

a lot of people here pointed out that prediction markets aren’t ...

i guess it’s about time for openai to drop a model ...

the obvious downside is that you won’t be appealing to every ...