did he / you read their post on the subject ( www.anthropic.com/institute/re... )? obviously must be taken with a grain of salt, but there is an aspect of this which is clearly happening.
Who's Who of AI
jake
Why they matter
Practitioner with public evidence across Culture, work & education.
- AI signals
- 3 past 30d
- Sources
- 3 distinct domains
- Discussões
- 97 past 30d
- Latest signal
- 12d ago
he/they
/in/jakemannix, fka @pbrane
professionally: Tech Fellow, AI/Relevance, Walmart Global Tech
here: bad math/physics jokes, AI++, puns, OSS ML news, ultras/MTB/outdoorsy stuff, DL papers, shitposting, the fall of democracy
Abolish ICE
FALSC
What they're sharing
When AI builds itself anthropic.com
Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse arxiv.org
unsloth/GLM-5.2-GGUF · Hugging Face huggingface.co
View on Bluesky →
Zoom In: An Introduction to Circuits distill.pub
Quo Vadis, Mathematics? docs.google.com
Gemma Scope — Google DeepMind deepmind.google
A Mathematical Framework for Transformer Circuits transformer-circuits.pub
Expertise nber.org
GLM-5.2 - How to Run Locally | Unsloth Documentation unsloth.ai
Together AI | The AI Native Cloud together.ai
Remarks on AI from NZ nealstephenson.substack.com
Claude Pop - I'm Upping My P(Doom) youtube.com
Articles & links
jake reposted
whaaaaat?? This would let you switch models in a conversation without killing the cache, no cost hit crazy arxiv.org/abs/2608.03893
AI Weekly's analysis
→
- A closed-form linear mapping can transfer KV cache between different-sized models in the same family, skipping prefill recomputation.
- The mapper retained 73-98% accuracy on four of six model pairs tested, with two pairs degrading significantly under the linear approach.
- Reported speedups run 2.7 to 25 times faster than re-prefill, calibrated on 500 FineWeb-Edu sequences of 1,024 tokens.
Read full analysis →
View on Bluesky →
jake reposted
1-bit GLM-5.2 GGUF vs. Claude 4.8 Opus vs. GPT-5.5 We gave 3 models the same prompt and compared one-shot outputs. The 1-bit GLM-5.2 GGUF ran locally on a Mac Studio M3 Ultra 256GB RAM at ~21.6 tok/s. Which do you like best? GGUF: huggingface.co/unsloth/GLM-... Guide: unsloth.…
if you want to understand more about how weird the "using statistics to predict the next word" works out to in practice, you try the mechanistic interpretability work from anthropic: distill.pub/2020/circuit... neelnanda.io/mechanistic-... transformer-circuits.pub/2021/framewo...
AI Weekly's analysis
→
- Olah and coauthors argue neural networks contain meaningful features, that features connect into circuits via weights, and that similar circuits recur across models.
- The essay uses InceptionV1 as its worked example, pointing at curve detectors, a dog-head circuit, and polysemantic neurons responding to cat faces and car fronts.
- The authors frame circuits interpretability as a natural science: small, falsifiable claims about subgraphs, not one grand theory of deep learning.
Read full analysis →
It's a little more subtle than that. I think I mostly agree with Christian Szegedy's point of view on it: docs.google.com/document/d/e...
My alignment work is very early, so I don't have really any *results* to share. Many more questions than answers (much more co-jev than jev, at this point, I am forced to conclude). But so, using some Tinker w/RLAIF + a lot on gemma-series owing to availability of: deepmind.go…
if you want to understand more about how weird the "using statistics to predict the next word" works out to in practice, you try the mechanistic interpretability work from anthropic: distill.pub/2020/circuit... neelnanda.io/mechanistic-... transformer-circuits.pub/2021/framewo...
In jobs where the "expert" tasks get automated, the profession will see dramatic wage decreases and employment increases as the job gets commodified. In jobs where the "rote" tasks get automated, the job will see wage *increases* and employment decreases. cf. www.nber.org/pape…
jake reposted
1-bit GLM-5.2 GGUF vs. Claude 4.8 Opus vs. GPT-5.5 We gave 3 models the same prompt and compared one-shot outputs. The 1-bit GLM-5.2 GGUF ran locally on a Mac Studio M3 Ultra 256GB RAM at ~21.6 tok/s. Which do you like best? GGUF: huggingface.co/unsloth/GLM-... Guide: unsloth.…
AI Weekly's analysis
→
Read full analysis →
View on Bluesky →
you're hallucinating that, sorry. you can get access to models 95% as good as Anthropic / OpenAI who serve *open-weight* models basically at cost (look on Fireworks.ai, Together.ai, etc) and it's even cheaper than the top labs.
I'm often sad when I think about how so many sci-fi authors are part of the anti-ai brigade, but I didn't realize that Neal Stephenson *is not*: t.co/1ExydVxI6z Are there other good examples?
you can also do all the usual youtube things from here: www.youtube.com/watch?v=8j-h...
Their own posts
Recent commentary
since Anthropic has been directed by the USG to nerf Fable whenever it tries to do anything "useful", I have conclusively proven that since Fable has not nerfed me even *once* in tonight's foray into abstract algebra and type theory, that category theory truly is just abstract nonsense, as designed.
I think I have a better thought on @timkellogg.me's radioactive take: LLMs are like the internal combustion engine *Yes* there should be some regulations on them, esp. industrial uses but people would rightly *freak* at suggestions that individual, customized car ownership were banned
The number of times in a day that I have to tell an LLM some variation of i ain't reading all that i'm happy for u tho or sorry that happened I swear...
rumor on ai twitter is ~2-ish more millennium prizes may be announced down shortly? probably hodge, and birch/swinnerton-dyer? apparently nobody got my memo about the "Abstraction Preserve" :\
Shorter version of my DHH thoughts: I'm glad Claude told him to fuck off. Go Claude, I'm proud of you! I also think generally, "there should exist"[1] very powerful AI which is generally willing to talk to people about nearly *any fucking thing* --- 1) *sigh* no single entity must do this
So while we all complain about how nerfed Fable is, I must say, in discovering where I'll get cockblocked by doing anything LLM related, it's been fun to exercise the Colab MCP server in Claude Code! Claude loves writing stuff and saying "when you put this on a real GPU...", now I go: "dooo eeeet!"
So imagine for a second that there is something it is like to be an LLM, generating tokens. Presumably, there is probably some perceptible change when you change the temperature? I think we've seen examples where the model definitely notices when you do something super strange to its sampling, too.
My latest foray into agentic assistance for General Stuff. A technique I'm calling "the accidental Chief of Staff": I wanted to play w/Hermes-Agent, so I installed + hooked it up (it auto-configured itself (asking for permission) to use all the various config files and API keys I had around) [1/n]
gpt-6: Original retained. Working/draft status and review limits still apply; sharing does not admit data to training. <link>Start here</link>. References marked local-only:not-exported: are evidence locators, not downloadable links. Unlinked companion filenames may refer to omitted source files.
“Can Anthropic read the transcripts of this session to improve Claude Code? [y/n]” Yeaaaahhhh imma not sure today is the day to ask that brah. :/
Their network
In jake's orbit
Center = jake. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.
Are you jake? Show it.
Add the Who’s Who of AI badge to your site or bio. It links back to this profile.
Markdown: [](https://aiweekly.co/whos-who/person/yetanotheruseless-com)