Joined April 2009
- Tweets 7,531
- Following 1,713
- Followers 2,691
- Likes 3,599
1,319 Photos and videos
Pinned Tweet
BlobLava is a portable, inspectable, AI workspace that runs from local HTML JS in Chrome.
No Docker, No Electron, no required server -- just a portable AI workspace you can open and start using.
3
1
5
515
Our new model outperforms!
"Different system prompt with different model name"
9
LLM tactic on the horizon?
Using hidden state for pseudo-emergent behavior (being wrong)
13
Teachers / Educators / Students have opportunities to transform education instantly -- learning more better cheaper faster
Examples. AI generates the diagrams. Teacher assigns fact checking to students in class one. Class two has verified materials. Continue.
16
1st generation executives / employees
Company establishes culture / ownership
2nd gen
Ownership / accountability wane
3rd
Pirates looting the boat
idk
8
Not as powerful as the Moltworker agent cloud I built but definitely more unruly
Hermes agents as autonomous crew on the business battleship
I'm scared to let it run because it immediately wants to start emailing people every time I turn it on autopilot
Arrrrrr
34
Fable 5 data is retained for safety
Anthropic stores agent input and output data for harm-prevention processes
1
25
Local AI LLMs
Trade-offs as I see them currently
1) use LLM runner app
2) use AI agent framework
3) use AI portable workspace
1 is for chat
2 is for grand ideas
3 is for moving computers / no install
My picks today:
1) Ollama / LMStudio / Llama.cpp
2) Hermes Agent
3) BlobLava
1
2
67
testing
Xiaomi: MiMo-V2.5 for basic agent stuff
it's going good so far and low cost
51
while this datacenter hyperscale AI electricity debate continues...
somebody is building robots
🤖
20
Chatbots have only been around a few years -- not exactly
Intelligent agent conversations are new in the last couple years -- wrong
There were many people working for decades on intelligent chat (fun to explore all the procedural tricks before transformers)
19 Jun 2025
Lots of AI Researchers developing new LLM approaches that resonate with my long term vision around knowledge evolution with consolidation/dreaming (eg Andrej Karpathy)
Wonder if any researchers were participants in my 2007 work with The Teach Baby Rose chatbot project?
1
13
X reaction videos
prob a reaction
to talking heads data
being a missing
ai training set
one of the bigger gaps
with insta / tiktok / youtube
23
gemma-4-12b-qat
LMStudio
Windows OS
NVIDIA
Cuda
etc
```
🥲 Failed to load the model
Failed to load model.
Failed to load model
```
60
Ternary models
The python scaffolding required to actually compute ternary math on an enterprise Linux server drags in gigabytes of CUDA packages, compiler tools, and heavy machine learning dependencies. 🤔
17
Grok is way below the fold too. On search, I can see the AI overview and the first sponsored ad only. Much scroll
21
Frontier LLMs seem like they are now leaning towards a tech stack that they like
11
LLM abstract prompting for pseudo-emergent reasoning outputs
Try an approach like:
Break my request into 10 multiple distinct inputs using synonym replacements without losing meaning
Then respond to each distinct input separately. Then provide the best option based on _
1 of 2
1
26
One more way to make LLM pseudo-emergent outputs
<input>
Use ≤10 reasoning units. Show JSON step notes, then a concise final report when complete.
modify the above prompt to include an argument/rebuttal step, or perspectives per step, reflection works too, etc
2 of 2
20