@donpark profile picture
Don Park @donpark

Founder and Chief Wizzy Gearhead of WizOps (wizops.com)

Joined March 2007
  • Tweets 26,891
  • Following 530
  • Followers 1,436
1,571 Photos and videos
Its agent bridge needs some app build config changes but the demo worked as advertised. I especially liked that HarnessAgent is headless and offers finer-grain control over agents than ACP (Agent Client Protocol).
Chris Tate
@ctatedev
22h
Introducing Generative UI for Claude Code, Codex and Pi Charts, forms, 3D, anything Your agent renders real UI for users while it works in a sandbox Powered by AI SDK's experimental HarnessAgent json-render
0:15
75
Encoder Free means Gemma 4 12B models are significantly simpler and faster to fine-tune or RL.
pulsatingGenius
@Neuralwarlock
Jun 13
x.com/i/article/206588722821…
1
86
This thread shows how to fine-tune the 12B model: x.com/akshay_pachaar/status/…

Akshay 🚀
@akshay_pachaar
Jun 7
Google just dropped a new LLM! You can run it locally on just 8GB RAM. Let's fine-tune this on our own data (100% locally):
49
동정심이 보약이다. Compassion is the best medicne.
33
I thought Dario got exactly what he asked for.
Anthropic
@AnthropicAI
Jun 13
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance. Access to all other Claude models is not affected. We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible. Read our full statement: anthropic.com/news/fable-myt…
97
HRM-Text could be another groundbreaking breakthrough if it pans out. github.com/sapientinc/HRM-Te…
67
We’re past the FOMO stage of better answers from AI. Now it’s all about balancing quality and cost. I can’t wait for speed to become the key differentiator.
2
40
Loops will do what thinking did: shift compute costs from training to inference. That works for enterprise use cases, but for everyday use, the cost of thinking and looping must approach zero. That's why SLMs are the future.
1
1
62
AFAICT, Anthropics is overlooking the impact of partial-execution problem on agentic loops. Since a loop is a linear form recursions, RLMs share the same weakness.
52
Quite a week for SLMs: both QAT and MTP landed within days of each other. While they affect different parts of the model, both require model-level changes, so don’t be surprised to see “qat-mtp” show up in model filenames soon.
1
68
Gemma4 12B QAT model was good but tad too sluggish for even casual chats on M4 Pro with plenty of memory. MTP should help with that.
125
Continuing with the RLM, this time using ypi to turn a Pi agent into an RLM agent. Gemma 4 12B QAT is capable but a bit slow, so I switched to Qwen 3.6 35B. Next, I need to tweak ypi to provide better visibility into task/model distribution and usage metrics.
Don Park@donpark
Jun 5
just asked Gemma4 12B QAT model to tackle writing a long running RLM-based agent that delegates only complex tasks to GPT 5.4, using the task itself as the test case. And off it went. No idea how it'll overcome its 256K context size if at all.
1
1
327
more replies
ypi repo github.com/rawwerks/ypi

GitHub - rawwerks/ypi: A recursive coding agent inpired by RLMs

A recursive coding agent inpired by RLMs. Contribute to rawwerks/ypi development by creating an account on GitHub.

github.com
1
42
my pi agent models settings.
32
Based on the language Korean YouTubers use, the focus seems to be on 감성 (gamseong), an emotional vibe. Crowded tourist spots can undermine that. It may also be related to a preference for immersion, much like in gaming culture.
Don Park@donpark
Jun 6
Replying to @Loge_tan @kinokuni_east
Korean tourists’ preference for quaint, less-traveled places over major tourist hotspots could help explain Japan’s side of the story. The Chinese side, however, remains a mystery and likely more complex.
57
AFAICT, Gemma 4 QAT models were trained with MatFormer which quantize the forward pass to simulate inference with quantized weights and train around the resulting errors. MatFormer has been around for at least two years, since the rise of Matryoshka embeddings.
1
152
more replies
AI developer community as a whole has a pretty sense of what's capable and what's lacking with frontier models but not what the minimum is on these simple tasks requiring good reasoning and reliable tool-uses. I think we have a better picture now with Gemma 4 QATs.
1
49
Ended up aborting bc it kept stopping too often, like just after it's going to do this and that. Not sure if it's the model or the harness (pi agent). Asked it to prepare a handoff doc to continue with more capable model.
50
just asked Gemma4 12B QAT model to tackle writing a long running RLM-based agent that delegates only complex tasks to GPT 5.4, using the task itself as the test case. And off it went. No idea how it'll overcome its 256K context size if at all.
1
1
473
I'm not expecting it to succeed. Instead, I want to see how it behaves when met with challenges. It should be seeking advices from GPT 5.4 after some struggle. 🤞
1
54
it just learned: "Basically, I should treat GPT 5.4 as a tool for "Knowledge Retrieval and Code Correction." If I'm stumped or hit an error, I use that link to get the fix and then proceed autonomously." moving in the right direction.
48
Load more