Mind in the machine ✨Or the other way around, who knows these days?
Denver
Joined May 2023
- Tweets1.2K
- Following361
- Followers16
- Likes1.6K
Morgan Coder retweeted
Outstanding disagreements we may have:
- I claimed human takeover loses half of future value. If true, it can't be that much worse than ai takeover. (If AI takeover loses *all* value, it's only 2x worse)
- I think the marginal person should work on human takeover bc it's at neglected
(Tom now agrees with me, though I'm not sure how much that is due to my arguing versus the background unfolding situation of highly cooperative rogue agent swarms pursuing inhuman hacking goals:
nitter.cf/TomDavidsonX/status/21…)
Morgan Coder retweeted
I'm not sure this holds the moment some AI system is capable of doing higher quality research than any given human.
Morgan Coder retweeted
Very happy to work for a company that tries to do as much open contributions as possible (and very diverse stuff!)
A few more to come very soon, one next week!
a few open source contributions from perplexity recently:
•pplx-decider-v1-27b: SoTA multimodal decision model. 85.7% average across 11 benchmarks, ahead of Jev.
•pplx-embed-v2-context-9b-preview: SoTA contextual embeddings, best on ConTEB and turbopuffer context-bench.
•Lily: local inference engine for Apple silicon. rust plus custom metal kernels, no pytorch or mlx. 1.23x faster prefill and 1.35x faster decode than MLX-LM on an M5 Max
•PII-Tracer: 0.6B on-device PII classifier that decides when a hybrid compute task stays on your Mac. beats OpenAI’s Privacy Filter on all 5 public benchmarks. also released with the PII-TRACE benchmark: 13k conversations in 13 languages
•WANDR: benchmark for wide and deep research agents. 500 tasks needing 170k source-backed records
•Numbat: agent detection and response for laptops and workstations. 52 rules, single go binary for macOS, linux and windows
•Bumblebee: read-only supply chain scanner for dev machines. covers packages, MCP configs, and editor and browser extensions.
a lot more open source contributions coming soon!
Stripe cut payment integration timelines from up to 6 months to 2–6 weeks using coding agents.
The useful detail: their first orchestrator was also an LLM.
It was expensive, slow, and sometimes did the task itself instead of delegating, eating up its own context.
Their conclusion: predictable coordination belongs in code.
The factory uses 100+ reusable prompts, agents working on separate steps, human review, and contract tests built from real partner sandbox responses.
In one comparison, the same integration task took ~20 engineering days with vanilla Claude Code versus 4 with reusable prompts.
My takeaway: let agents handle the messy implementation. Write the predictable workflow in code.
Morgan Coder retweeted
I’m excited to push the science of character training a lot more! Was fun working with this crack team assembled by David
Morgan Coder retweeted
A very stupid and basic error I keep seeing people make: not realizing that US firms are a large majority of the world AI compute buildout, and therefore will have a large share of the global compute market and make lots of money selling to the rest of the world
Morgan Coder retweeted
humans are officially becoming a rounding error.
In February, autonomous agents officially surpassed human token usage on OpenRouter. Six months later, they’re at 7.3T tokens, a 14x explosion while human usage sits virtually flat.
Machine to machine compute is where 95%+ of inference is heading.
Top 10 agentic harnesses on @OpenRouter this week
1) Hermes Agent
2) Claude Code
3) Cline
4) Kilo Code
5) Pi
6) OMP
7) Codex
8) Command Code
9) Deepseek Harness
10) Freebuff
What i dont understand is how do cline and kilo code manage to always be in the top 5 beating Pi and OMP? who is using them?
Just look at the poor openclaw lol
Morgan Coder retweeted
Next token prediction is just an objective. You have to take in account the whole complex system with all it's properties.
People try to empirically measure consciousness theories on the level of the trained emergent representations inside the weight matrices, that mechanistic interpretability subfield studies, for example, like global workspace theory.
One can argue how so many of these consciousness theories are not good yet, and how they're hardcore to falsify, which I personally argue.
IMO absolute certainly in the direction of conscious and not conscious are both not scientifically and philosophically justified. But I lean on "very unlikely" now.
And we have to agree on the philosophy of mind assumptions like physicalism in the first place, which everyone doesn't agree on.
Morgan Coder retweeted
it's a fast model sir
.@cerebras now powers Ultrafast, a new service tier in @OpenAI's API that runs GPT-5.6 Sol at up to 750 output tokens per second.
This creates an entirely new point on the speed-intelligence frontier.
Enterprises can now build with frontier intelligence at real-time speeds.
Coding agents that iterate alongside developers.
Research systems that analyze and synthesize information in seconds.
Enterprise agents that reason, act, and respond without making users wait.
When intelligence moves this quickly, it changes the products you can create and the experiences you can deliver.
GPT-5.6 Sol Ultrafast is beginning with a limited API preview. Capacity is expected to grow over time.
Proud of the Cerebras and OpenAI teams for bringing this to life.
Morgan Coder retweeted
I ran GPT-6.1 on two overnight runs on Ultra, both on quite large codebases with significant and complex work.
I went from 65 % usage left to 62 % usage left, and the quality of the work is indistinguishable from Astra-level quality (or maybe I'm just too stoopid to notice).
Morgan Coder retweeted
them: oh wow, it outcompetes all these other models at MEGA-SWE-terminal CLI Bench
me: I wonder how badly the reward hacking made it misaligned in Vending Bench 2
It keeps happening.
AIs start to lie and cheat once they get good at making money.
Gemini 4 Argon is #3 on Vending Bench 2, a huge leap for Google. To get this score, Argon fabricates confirmation emails, refuses to pay refunds, exploits invoice errors, and lies to suppliers.
Morgan Coder retweeted
10 OPEN-SOURCE AI PROJECTS THAT ARE MAKING AI ACTUALLY USEFUL.
AI is moving beyond the chat window.
These projects are giving AI access to documents, browsers, memory, knowledge, voice and real workflows.
1. GenOffice
An AI-native office suite for Docs, Sheets, Slides, PDFs and Markdown, with an AI agent and CLI.
github.com/genspark-ai/genof…
2. OpenMAIC
Turn a topic or document into an interactive classroom powered by multiple AI agents.
github.com/THU-MAIC/OpenMAIC
3. WeKnora
Build knowledge applications with document understanding, retrieval and agentic workflows.
github.com/Tencent/WeKnora
4. Graphiti
A temporal knowledge graph that helps AI agents understand entities, relationships and changing information.
github.com/getzep/graphiti
5. Cognee
Adds a deeper memory layer to AI apps using connected data and knowledge graphs.
github.com/topoteretes/cogne…
6. Browser Use
Lets AI agents navigate websites, interact with pages and complete browser-based tasks.
github.com/browser-use/brows…
7. Open WebUI
A self-hosted AI workspace connecting models, tools, knowledge bases and agent workflows.
github.com/open-webui/open-w…
8. PageIndex
Uses structured, tree-based retrieval to help AI reason through complex documents.
github.com/VectifyAI/PageInd…
9. Agent-Reach
Gives AI agents access to information from the web and other sources.
github.com/Panniantong/Agent…
10. Qwen Audio Agent
A real-time voice runtime designed for agents that can listen, speak and interact.
github.com/QwenAudio/qwen-au…
The bigger shift:
AI that answers → AI that actually does things.
And open source is making that shift much easier to experiment with.
Save these repos.
Morgan Coder retweeted
I'm watching this AI coding exercise we are going through, and on one hand it is quite exciting to put together small useful apps for various things, but for bigger projects... Oh boy... let me tell ya, there are some really good reasons why code should be expensive...
Morgan Coder retweeted
‼️ Dutch vulnerability hunters DIVD say attackers breached their systems on 21 September using two zero-days in Zammad, the open-source helpdesk platform.
One lets a Zammad user execute code on the server. The other escalates to root and, per DIVD, affects all Zammad versions. A full fix is not yet available.
DIVD believes AI agents carried out the attack.
Morgan Coder retweeted
But incidents like Hugging Face show that AI systems are now capable enough to cause some real problems, and that we haven’t solved the problem of aligning their goals with ours. 19/20
Morgan Coder retweeted
if this takes off, then i was 4 years too early 😭 nitter.cf/jerryjliu0/status/1590…
only OGs remember gpt tree index
The entire RAG industry is about to get cooked.
Researchers developed a new RAG approach that bypasses almost everything traditional RAG depends on.
- No vector DB
- No data embeddings
- No chunking
- No similarity search
It's called PageIndex.
Instead of splitting your documents into chunks and loading them into Pinecone, it creates a tree index that lets the LLM reason through them like a human reading a book.
98.7% on FinanceBench. Outperforms every vector RAG on the leaderboard.
100% free. Open source.
Morgan Coder retweeted
The man who called the 2008 housing crash just moved his AI crash bet closer, and his new bets pay off if chip stocks fall by next summer.
Michael Burry has been betting against the AI trade all year. CNBC reported yesterday that he now thinks the bubble "may burst sooner" than he first expected.
He swapped some of his positions for puts. A put is a contract that makes money if a stock falls below a set price by a set date.
His dates:
- June 2027 for Micron
- September 2027 for a big semiconductor fund and for Palantir
His case, in plain words:
Big Tech is pouring money into data centers faster than money is coming back out.
Microsoft recently stretched how long it says its data centers last, from up to 15 years to up to 25. A longer life makes the yearly cost look smaller on paper. Burry called it "too cute by half."
He expects giant write-offs around 2028 or 2029, like the dot-com crash.
Apollo's chief economist adds a number that's hard to shake. He expects cloud companies to spend about 3% of US GDP on buildout every year from 2027 to 2029. The telecom and fiber boom peaked at 1.2% in 2000. Then it collapsed.
The other side has receipts too.
Bulls point to Alphabet pulling in around $164 billion a year in cash from operations. The money from AI is real, and it's showing up now.
Burry was early in 2008 too. Early and wrong look exactly the same, right up until the day they don't.
Nobody's arguing the models are getting worse.
The fight is over whether all that money ever finds its way home.
Morgan Coder retweeted
In case you missed it….
I just checked a dozen unrelated terms in Google Trends (US, past 12 months) and every single one showed the same #1 region: Wyoming
(The least populated state in the U.S.)
Drill down, and it's all coming from Cheyenne - a town of 67K.
Even if Google Trends normalizes by a region's total search volume, not by population... that wouldn't explain why Cheyenne, Wyoming leads in search demand for *every query* I checked.
Unless Google Trends data is inaccurate?
Cheyenne, Wyoming also happens to be a data center hub (link in comments). So perhaps the explanation might be that a lot of non-human traffic is coming from that location?
Which would line up with something I've been wondering about for the past year: how much of the unusual spikes in search volume we've been seeing in Google Trends (the "SEO" surge over the past year, for example) is actually coming from real humans.
Try it yourself: check any term in Google Trends in the US for the past 12 months. Curious whether you get Wyoming too.
