arthurcolle

> Of course, M1-M3 have broken secure enclave which cannot be fixed by the user.

haven't heard about this, link?

this could legit be pre-installed on a nokia phone in some alternate potential future!

great work

followup: the little black outlines in the video correspond to helicopter-like objects. I just referred to Apache for some reason, most local memory of an attack helicopter. apparently it's something else

TikTok videos showing Apache helicopters shooting missiles at targets. Lots of planes and helicopters flying over Caracas

Agent hacked the harness

  Location: Washington, D.C. (open to relocation, but preference to initial remote period)
  Remote: Yes
  Willing to relocate: Yes (June 2026 and beyond)
  Technologies: Python, Rust, Elixir | Neo4j, Postgres, Redis, K8s | PyTorch, MLX, LoRA/QLoRA, Modal A100s, H100s | FastAPI, Phoenix, Rails
  Résumé/CV: https://linkedin.com/in/arthurcolle
  Email: arthur@distributed.systems

Background:

Deep knowledge of OpenAI, Anthropic + LLM APIs. Able to run models on GPUs and Apple Silicon unified memory. I have trained + deployed agentic coding, tool calling and multi-turn agent models in two prior roles.

Building autonomous agent infrastructure. Currently run a small R&D shop (Distributed, Systems & Co.) doing distributed AGI and AI safety research. I am developing graphsub (graph substrate), a graph-native service mesh for agents, eval harnesses, self-modifying DSLs.

Before that: 4 years at Goldman Sachs deep in SecDB/Slang (the reactive dependency-graph system). I worked on collateralized mortgage obligations, on the interest rate products trading desk (as an engineer in the technology division), automated parts of the downstream IR-swap lifecycle and on CDS lifecycle management for novations and other lifecycle events. Had a hand in managing and working on our internal TriOptima notional compression cycles. I got promoted from tech into structuring. Series 7.

Recent work: 15-service LLM stack (zero downtime across 98 deploys), fine-tuned Llama-3.1-8B (23% hallucination reduction), Rust daemons and TUIs. 3k+ GitHub contributions last year.

OSS: agents.erl (BEAM-native cognition), https://github.com/arthurcolle/openai-mcp, fortitude (a framework for bootstrapping agents), https://github.com/arthurcolle/dspy.ex

Looking for: AI labs, defense tech, or startups building agentic systems / foundation model infra / AI safety tooling. IC or research roles.

I can run SFT and RL experiments, train models, and I'd love to be able to train agentic models and ideally with large GPU clusters. I have enjoyed these last two years, a focused period of building "agentic scaffolding" and scaling up multi-agent systems and if any of this sounds appealing to you, I can definitely be of some help to your org!

GitHub: https://github.com/arthurcolle

recent research publications:

https://apartresearch.com/project/detecting-piecewise-cyber-...

Terrible idea. Will lead to nothing but chaos

Merry Christmas everyone. It was a pleasure interacting with you all this year and I hope for many more.

Checks and balances for a more technological era.

Also good for UFO/UAP/"anomalous phenomena" documents and remote viewing PDFs for what it's worth :)

check out My Tank Is Fight! if you haven't

so many weird weapons developed over the years

Why would you pay for retrieval when you could pay for completed correct answers?

Wrong to the point of being misleading. This is a goal, not an assumption

Source: all of mechinterp

That would suck. Is Google going to just eat all of this?

That's too easy. It'll be more subtle. Compatibility MCP-Gemini for "security" so it slurps in more data from all the other AIs

Meta-vibing. I like it

Thank you for your concern, but Anthropic researchers themselves describe their misaligned models as "evil" and laugh about it on YouTube videos accessible to anyone, such as yourself, with just a few searches and clicks. "We realized the models were evil" is a key quote you can use to find the YouTube video in the transcripts from in the past two weeks.

I didn't think the language in the post required all that much imagination, but thanks for sharing your opinion on this matter, it is valued.

It has been noted, by whom? Their system cards?

It is deprecated and unavailable now, so it's convenient that no one has the ability to test these theses any longer.

In any case, it doesn't matter, this was over a year ago, so current models don't suffer from the exact same problems described above, if you consider them problems.

I am not probing models with jailbreaks making them behave in strange ways. This was purely from a eval environment I composed where it is asked to repeatedly asked to interact with itself and they both had basically terminal emulators and access to a scaffold to make them able to look at their own current 2D grid state (like a CLI you could write yourself and easily scroll up to review previous AI-generated outputs)

These child / neighbor comments suggesting that interacting with LLMs and equivalent compound AI systems adversarially or not might be indicative of LLM psychosis are fairly reductive & childish at best

Fair enough, thanks for your insightful comment.

I personally am fairly convinced that there is emergent misalignment in a lot of these cases. I study this and Claude 3 Opus was extremely misaligned. It would emit <rage> tags, and emit character control sequences if it felt like it was in a terminal environment, and would retroactively delete tokens from your stream, and all kinds of funny stuff. It was already really smart, and for example if it knew the size of your terminal shell, it would properly calculate how to delete back up to the positional cursor index 0,0 and start rewriting things to "hide" what it was initially emitting

I love to use these advanced models but these horror stories are not surprising

gpt-4-32k pricing was originally $60.00 / $120.00.

  Location: Washington, D.C. (open to relocation, but preference to initial remote period)
  Remote: Yes
  Willing to relocate: Yes (June 2026 and beyond)
  Technologies: Python, Rust, Elixir | Neo4j, Postgres, Redis, K8s | PyTorch, MLX, LoRA/QLoRA, Modal A100s, H100s | FastAPI, Phoenix, Rails
  Résumé/CV: https://linkedin.com/in/arthurcolle
  Email: arthur@distributed.systems

Background:

Deep knowledge of OpenAI, Anthropic + LLM APIs. Able to run models on GPUs and Apple Silicon unified memory. I have trained + deployed agentic coding, tool calling and multi-turn agent models in two prior roles.

Building autonomous agent infrastructure. Currently run a small R&D shop (Distributed, Systems & Co.) doing distributed AGI and AI safety research. I am developing graphsub (graph substrate), a graph-native service mesh for agents, eval harnesses, self-modifying DSLs.

Before that: 4 years at Goldman Sachs deep in SecDB/Slang (the reactive dependency-graph system). I worked on collateralized mortgage obligations, on the interest rate products trading desk (as an engineer in the technology division), automated parts of the downstream IR-swap lifecycle and on CDS lifecycle management for novations and other lifecycle events. Had a hand in managing and working on our internal TriOptima notional compression cycles. I got promoted from tech into structuring. Series 7.

Recent work: 15-service LLM stack (zero downtime across 98 deploys), fine-tuned Llama-3.1-8B (23% hallucination reduction), Rust daemons and TUIs. 3k+ GitHub contributions last year.

OSS: agents.erl (BEAM-native cognition), https://github.com/arthurcolle/openai-mcp, fortitude (a framework for bootstrapping agents), https://github.com/arthurcolle/dspy.ex

Looking for: AI labs, defense tech, or startups building agentic systems / foundation model infra / AI safety tooling. IC or research roles.

I can run SFT and RL experiments, train models, and I'd love to be able to train agentic models and ideally with large GPU clusters. I have enjoyed these last two years, a focused period of building "agentic scaffolding" and scaling up multi-agent systems and if any of this sounds appealing to you, I can definitely be of some help to your org!

GitHub: https://github.com/arthurcolle

recent research publications:

https://apartresearch.com/project/detecting-piecewise-cyber-...

I don't see a contact email listed, but would be happy to chat!

Couple days or weeks usually. No one is doing 9 month training runs

Surely OpenAI will follow up with a gpt-oss-780b

might want to give claude-opus-4.5 a spin if you haven't

Why isn't there a "if I like you you can use it"

I like you @WalterBright you can use any of my stuff even if you get acquired

You can feed a rope out of something (see: 3D printer extruders)

The behavioral modeling is the product

Classified