iiinova — where your digital self keeps evolving

My AI agent dev workflow: I start by writing a design document for a new feature or a refactor of an existing part of the codebase. I review and revise that document repeatedly, often around 10 times, until I can no longer identify any significant issues or ambiguities.

192

Scaling Pre-Training in Practice: A Hierarchical Approach by Jordan Sassoon It’ll guide you through the choices we make at different GPU counts, taking a 30B-A3B MoE model from 16 - 512 B200 GPUs with 35% MFU & near-linear scaling. aleph-alpha.com/en/blog/scal...

190

Learning from the 7,780 environments Xiaomi open-sourced for MiMo. In RL environments, reward design is everything huggingface.co/spaces/FineE...

0224

Reshaping Monte-Carlo Tree Search 2FFS A new tree search algorithm that combines multi-fidelity bandits to resolve the fundamental trade-off: should we use cheaper, approximated evaluations, or expansive but accurate samplings? Paper: arxiv.org/abs/2606.01708

0191

The Bitter Lesson for context management: Giving LMs unrestricted control over their context beats human-designed SOTA! - Natively manage their own context - Treat context as a file - Learn policies in CLM weights, no harness

2554

Tokenization: A Survey for Modern NLP Over the past ~8 months, 32 (!) tokenizer researchers put together the comprehensive survey of the field. www.alphaxiv.org/abs/2609.tok...

0183

Word in the street on gpt-6.1-sol being slow....

1160

Christopher Degnan of Snowflake announces that he is joining Cognition Labs. Matan Grinberg of Factory AI accuses Degnan, who had served as an advisor to Factory AI, of stealing confidential information and taking it to Cognition Labs.

100

I really miss the old days when one of the most exciting things in machine learning was XGBoost.

4722

We're back to three way race again.

3212

Cube: an always-on cloud computer for Claude Code & Codex

Is anyone using cloud 'dev' computer like Cube Computer (cube.computer) or boat (boat.dev) for your development as well as your agents? They seem affordable.

cube.computer

Run coding agents on your Mac or a personal cloud computer. Keep projects, terminals, and previews together in Cube.

271

It's a bit funny that OpenAI's Tibo went from one of the most beloved people to one of the most disliked people in a day. Now, they're openly fighting. 🤣🤣🤣

020

I've been developing with OpenAI's Dots since yesterday, and here's my take: 1. When developing with Codex, it provides constant feedback on what it is working on. Dots does not provide any feedback unless asked.

240

I'm so confused. What is Codex task in this context?

310

In the old days, young people entering the workforce had a technological advantage because they had grown up with tools that older generations had to learn later in life. But AI is relatively new to everyone, and judging its performance requires knowledge of the business domain.

2162

Dots by OpenAI are proactive assistants that can keep working across complex projects and everyday tasks. Learn how dots help you stay in control while work moves forward.

My review of OpenAI dots (openai.com/index/introd...). I'm using it as a coding agent. It seems to be working well since it uses gpt-6-astra and its usage is NOT counted against my usage. I'll be using this a lot while it's kind of free.

openai.com

Introducing dots

2192

How about creating a decision model or Yet Anther Jev, what if someone create a LLM-based decision tree model. Just saying... Input: { "question": "Does this PR require security review?", "context": "...", "branches": [ "security_review", "normal_review", "skip_review" ] }

360

Dots: Always-on agents built to handle everything | ChatGPT

It feels like OpenAI developed dots (chatgpt.com/features/dots) to counter Claude tag (www.anthropic.com/news/introdu...), but made it cute to counter Meta's Muse (ai.meta.com/muse/).

chatgpt.com

Meet your dot. It learns your priorities, takes on ongoing tasks, and keeps making progress between conversations, giving you more time for what matters.

0140

Nvidia's Kumo Tabular, a new family of foundation models for tabular data - open weights, open-source software, and a permissive license for commercial use. HuggingFace: huggingface.co/nvidia/Kumo-... GitHub: github.com/NVIDIA/struc...

2322

Jev's moat lasted like a day. LiquidAI announces d1, their decision model. > Decision models page: docs.liquid.ai/lfm/models/d... > Migration guide: docs.liquid.ai/guides/decis... > Demo tutorial: github.com/Liquid4All/c...

291

OpenAI dots is FREE for a limited time. Dots uses gpt-6-astra I can create unlimited dots. I'm just using dots for all my software development.

2141

OpenAI: we're aligning our API pricing with subscription pricing. Now, you can just use your subscription with third-party apps. Thanks, I guess...

1210

claude tag meta muse openai dot

020

There probably will be a codex reset tomorrow, so you may want to drain your usage.

3160

Is your GitHub Actions' cost hitting 3 or 4 figures? May I recommend self-hosting your own Git Runners?

180

Sung Kim @sungkim.bsky.social · 28/09/2026

My thoughts on Instinct AI: It's too cool for its own good and it does not support Android's SMS/RCS.

010

Google's RRSI: Regularized Recursive Self-Improvement of Agent Harnesses github.com/google-resea...

0232

Sung Kim @sungkim.bsky.social · 28/09/2026

I don’t know about you, but I trust Meta more than some startup with a pretty UI wrapped around an AI agent. Time and time again, we find out that these products launched with serious security gaps that should have been addressed from day one. Wajo wajo.ai/join-wajo

wajo.ai

Join Wajo

Give Fo the things you need done. Fo calls, emails, books and pays – then follows through until the job is finished.

340

Reposted by Sung Kim

mr. TIM @timkellogg.me · 28/09/2026

Claude 5.5 Sonnet is live and it’s roughly Opus 5.5 but cheaper www.anthropic.com/claude-sonne...

A benchmark comparison table titled "Claude Sonnet 5.5" comparing four models: Sonnet 5.5, Sonnet 5, Opus 5.5, and GPT-6 Sol.
* Agentic coding (Terminal-Bench 4.0): Sonnet 5.5: 70.6%; Sonnet 5: 10.3%; Opus 5.5: 66.4%; GPT-6 Sol: N/A.
* Agentic coding (FrontierCode 1.1 Main): Sonnet 5.5: 46.2% (Max) / 52.1% (Xhigh); Sonnet 5: 42.4%; Opus 5.5: 54.4%; GPT-6 Sol: 49.3%.
* Agentic coding (CursorBench 4.0): Sonnet 5.5: 55.5%; Sonnet 5: 34.1%; Opus 5.5: 57.8%; GPT-6 Sol: N/A.
* Knowledge work (GDPval-AA v2.1): Sonnet 5.5: 1844; Sonnet 5: 1449; Opus 5.5: 1846; GPT-6 Sol: 1487.
* Knowledge work (AA-Briefcase v1.1): Sonnet 5.5: 1811; Sonnet 5: 1359; Opus 5.5: 1822; GPT-6 Sol: 1483.
* Multidisciplinary reasoning (Humanity's Last Exam with tools): Sonnet 5.5: 64.5%; Sonnet 5: 54.9%; Opus 5.5: 67.7%; GPT-6 Sol: N/A.
* Computer use (OSWorld 2.1 partial): Sonnet 5.5: 80.1%; Sonnet 5: 57.0%; Opus 5.5: 81.8%; GPT-6 Sol: N/A.
* Visual chart recognition (Chartography no tools): Sonnet 5.5: 61.6%; Sonnet 5: 15.6%; Opus 5.5: 64.4%; GPT-6 Sol: 53.6%.
Footnotes provide methodology notes regarding evaluation settings, Artificial Analysis pre-release testing details, and recent bug fixes affecting GPT-6 Sol benchmark scores.

A line graph titled "Agentic coding by effort level" on the CursorBench 4.0 benchmark, plotting Score (%) on the linear y-axis (20% to 60%) against Cost per task in USD on a logarithmic x-axis ($0.5 to $10+).
The chart compares four models:
* Sonnet 5.5 (blue line with labeled effort levels): Starts at "Low" (~$0.50, 36%), moving through "Med" ($0.70, 39%), "High" ($1.70, 48%), "Xhigh" ($3.80, 53%), and "Max" ($9.50, ~55.5%).
* Opus 5.5 (orange line): Tracks closely above Sonnet 5.5 at higher cost points, spanning from ~$1.20 per task (~44%) up to ~$12.50 per task (~58%).
* Sonnet 5 (grey line): Shows lower accuracy relative to cost, ranging from ~$0.90 per task (~25%) to ~$8.00 per task (~42%).
* GPT-5.6 Sol (light green line): Represents the lowest trajectory, ranging from ~$1.40 per task (~24%) to ~$7.00 per task (~34%).
Footnote: "CursorBench 4.0 does not report GPT-6 Sol performance publicly, so we report GPT-5.6 Sol here."

2020623

Sung Kim @sungkim.bsky.social · 28/09/2026

From my estimation, gpt-6-astra runs for about 2 days on a 20x plan and claude-opus-5.5 runs for about 2.5 days on a 20x plan. I do not hit 5 hour limit on claude-opus-5.5.

2130

So... I’ve automated the code review process for my personal projects, and it’s actually getting kind of boring now that agentic AI coding harnesses can run for days at a time.

360

MiniMax kind of launched M3.1-Flash-Preview, only for MiniMax Code users.

0120

If you're buying an iPhone 18 Pro in the U.S., consider the iPhone 18 Pro Max instead of the regular Pro. It's all about the cellular modem. The U.S. iPhone 18 Pro Max uses Qualcomm's modem, while the iPhone 18 Pro uses Apple's new C2 modem.

5141