Brandon Ong

291 posts

Brandon Ong banner
Brandon Ong

Brandon Ong

@bytedunks

exploring; prev Joint PhD in Robotics (on leave) @Columbia @NTUsg; @join_ef

SG Katılım Mayıs 2018
261 Takip Edilen413 Takipçiler
Sabitlenmiş Tweet
Brandon Ong
Brandon Ong@bytedunks·
To understand what it takes to build a humanoid robot with model-based control, we finetuned @physical_int 's (PI) Pi05 model for our custom use case and environment. We incurred ~$10K in hardware costs, compared to the typical ~$20K set up (DROID/ALOHA). Here are the lessons and challenges we faced building the first working prototype (shown in the video) in 3 months. Part 1: Hardware, Software, Model Selection, Custom Embodiment, Inference, Embedded Hardware, Hierarchical Planner Part 2: Model Evaluation, Data Collection, Model Training, Simulation and Teleoperation We hope sharing our experience accelerates the learning of others who are in a similar starting point.
Brandon Ong@bytedunks

x.com/i/article/2018…

English
3
11
110
16.2K
Brandon Ong retweetledi
General Reasoning
General Reasoning@GenReasoning·
Introducing OpenReward. 🌍 330+ RL environments through one API ⚡ Autoscaled sandbox compute 🍒 4.5M+ unique RL tasks 🚂 Works like magic with Tinker, Miles, Slime Link and thread below.
General Reasoning tweet media
English
16
137
1K
144K
Brandon Ong
Brandon Ong@bytedunks·
touche, applying world models to codegen - excited to try if you have a consumer product congrats on the launch
Animesh Koratana@akoratana

Introducing: PlayerZero The world's first Engineering World Model that puts debugging, fixing, and testing your code on autopilot. We've raised $20M from Foundation Capital, @matei_zaharia (Databricks), @pbailis (Workday), @rauchg (Vercel), @zoink (Figma), @drewhouston (Dropbox), and more PlayerZero frees up 30% of your engineering bandwidth by: 1.⁠ ⁠Finding the root cause for bugs & incidents in minutes that engineering teams take days to identify. 2.⁠ ⁠Predicting in minutes, edge case issues that a 300-person QA team would take weeks to find. ------ Here's why this matters: No one in your org has a complete picture of how your production software actually behaves. Support sees tickets. SRE sees infra. Dev sees code. Each team builds their own fragmented view - and none of these systems talk to each other. When something breaks, everyone scrambles to stitch the picture together by hand. PlayerZero connects all of it into a single context graph - → The Slack thread where your lead said "we went with X because Y fell apart in prod last time" → The PR review where an engineer explained the tradeoff → The lifetime history of your CI/CD pipeline, observability stack, incidents, and support tickets So you can trace any problem to its root cause across every silo. And it compounds. Every incident diagnosed teaches the model something new. The longer it runs, the deeper it understands - which code paths are high-risk, which configurations are fragile, which changes tend to break which customer flows. So when you sit down to debug a live issue, you have your entire org's collective reasoning and production memory behind you - instantly. ------ Zuora, Georgia-Pacific, and Nylas have reduced resolution time by 90% and caught 95% of breaking changes and freeing an average of $30M in engineering bandwidth. ------ Our guarantee: If we can't increase your engineering bandwidth by at least 20% within one week, we'll donate $10,000 to an open-source project of your choice. Book a demo - bit.ly/3NlLMeN

English
0
0
1
182
Brandon Ong retweetledi
Jenny Zhang
Jenny Zhang@jennyzhangzt·
Introducing Hyperagents: an AI system that not only improves at solving tasks, but also improves how it improves itself. The Darwin Gödel Machine (DGM) demonstrated that open-ended self-improvement is possible by iteratively generating and evaluating improved agents, yet it relies on a key assumption: that improvements in task performance (e.g., coding ability) translate into improvements in the self-improvement process itself. This alignment holds in coding, where both evaluation and modification are expressed in the same domain, but breaks down more generally. As a result, prior systems remain constrained by fixed, handcrafted meta-level procedures that do not themselves evolve. We introduce Hyperagents – self-referential agents that can modify both their task-solving behavior and the process that generates future improvements. This enables what we call metacognitive self-modification: learning not just to perform better, but to improve at improving. We instantiate this framework as DGM-Hyperagents (DGM-H), an extension of the DGM in which both task-solving behavior and the self-improvement procedure are editable and subject to evolution. Across diverse domains (coding, paper review, robotics reward design, and Olympiad-level math solution grading), hyperagents enable continuous performance improvements over time and outperform baselines without self-improvement or open-ended exploration, as well as prior self-improving systems (including DGM). DGM-H also improves the process by which new agents are generated (e.g. persistent memory, performance tracking), and these meta-level improvements transfer across domains and accumulate across runs. This work was done during my internship at Meta (@AIatMeta), in collaboration with Bingchen Zhao (@BingchenZhao), Wannan Yang (@winnieyangwn), Jakob Foerster (@j_foerst), Jeff Clune (@jeffclune), Minqi Jiang (@MinqiJiang), Sam Devlin (@smdvln), and Tatiana Shavrina (@rybolos).
Jenny Zhang tweet media
English
131
543
3K
265.4K
Ben Lang
Ben Lang@benln·
We're taking over a cafe on March 30th in Singapore Grab coffee, Cursor credits, meet the team, and build together
Ben Lang tweet media
English
43
15
408
25.1K
kingston kuan
kingston kuan@kstonekuan·
I used Gemini CLI as an agent harness with the new Google Workspace CLI to generate docs, slides, and sheets for exploring new ideas
English
4
1
6
334
Brandon Ong
Brandon Ong@bytedunks·
@ivanleomk Could be the reason behind the influx of GDM reposts ;)
English
1
0
1
30
Ivan Leo
Ivan Leo@ivanleomk·
Big changes in the next 2 weeks
English
9
0
45
2.7K
Brandon Ong
Brandon Ong@bytedunks·
GitHub: github.com/theogbrand/sel… I wanted to try Gemini 3.1 Pro, which is known for relative cost effectiveness and inference speed. This is true in practice. However, I faced an unexpected issue with API performance degradation: During the development of this project, the model performance via Gemini API abruptly degraded, and my agentic tool calling capabilities completely broke when this happened. I faced this performance degradation consistently over 2 weeks at around 5-9PM SGT. From my understanding, something similar to this was previously reported with Anthropic APIs and could be a known inference issue that is being addressed. Running inference at low latency for multi-step tool calling for agentic workflows is understandably challenging. I will be curious if others have a similar experience with the Gemini API or others. Will be integrating Anthropic and OpenAI next.
English
0
1
1
107
Brandon Ong
Brandon Ong@bytedunks·
Lately, I have been thinking about continual learning --- could this unlock novel interfaces to solve more complex problems with AI? An experiment: Self-Improving Browser Agents (SIBA) Browser agents are currently difficult to steer just by prompting, making them hardly capable at executing workflows reliably and flexibly without parametric updates. I could never get actual tasks done with Browser Agents previously, and thought this could be a worthy attempt at changing this. The demo shows the SIBA agent extracting 3 receipts from my Gmail, and saving them into a local directory, tagged by software, for what would be a monthly software claim submission process. What differentiates SIBA from previous methods: 1) CLI-driven browser control Following agent-browser, we select web components by textual references which significantly reduces incorrect component selection. Furthermore, LLMs work best reasoning about code. Previous methods rely heavily on a model's visual understanding to select the correct visual component by generating the (x,y) coordinates of a browser window (which are expected to be almost precise). The emerging "Skills + CLI" design pattern for agentic tool calling will be included in the next iteration. 2) Meta-agent driven improvement Tuning a browser agent's system by hand is ineffective: the root cause of most failure modes are ambiguous and difficult to address systematically without parametric updates. Meta-agent driven improvement based on human feedback show promising initial results for now --- it took me three iterations of feedback and some API credits to get my custom workflow working. Moreover, providing clear feedback is something humans are good at, and also a user experience that feels natural. Caveat: currently, a simple meta-agent is used for iterative improvement. There are more effective ways to do so systematically that will be explored next. My hope is to use SIBA for smoke testing at scale. It would be interesting to see if it could be extended to persona-based smoke testing using parallel agents. Let's see.
English
1
1
6
685
Brandon Ong
Brandon Ong@bytedunks·
Not bad for a fishing village with no natural resources (Singapore) that 180x’d its GDP per capita from $500 to $90,000 in 60 years.
Olivia Moore@omooretweets

Our team @a16z calculated AI adoption per capita across the world. The results were surprising. The U.S. leads AI development...but it ranks down at #20 in adoption. At the top? Singapore, Hong Kong, the UAE, South Korea, and much of Europe 🤯

English
0
0
1
113
Brandon Ong
Brandon Ong@bytedunks·
@sooneggg @trq212 ah MCP is a good idea. Somehow CLIs are also blocked by egress policies in Claude Cowork. I faced issues with Coingecko MCP somehow it doesn't work, and defaulted to Crypto.com MCP as recommended by Claude
English
1
0
0
46
Brandon Ong
Brandon Ong@bytedunks·
hey @trq212 what is Anthropic's thinking behind disallowing users from allowlisting their own sites if code is ran in a sandbox anyway --- or is this a feature coming soon? I am trying to set up a scheduled task in CoWork to monitor BTC prices from Coingecko's public API but currently am only able to do this through Web Search.
English
1
0
2
88
Brandon Ong
Brandon Ong@bytedunks·
turns out you can build a school for builders
Alice Bentinck@Alicebentinck

Today, we're announcing @join_EF's Series D: $200M of fresh capital, at a unicorn+ valuation. We're in a golden age of entrepreneurship. More individuals than ever before are taking the leap, with the world's most transformative technology at their fingertips. For over a decade, we've backed extraordinary people without asking what idea they're working on. We call this Talent Investing; believing in someone from Day 1 and giving them the peer group, environment, and capital to find their life's work. And it works! Companies built through EF are now worth $16bn, up from $3bn at our last raise in 2021. Thank you to our investors for their continued belief, the EF team for bringing this vision to life, and the founders who rejected the status quo to build globally important companies with us. And, of course, a huge thank you to @matthewclifford, my exceptional cofounder.

English
0
0
2
119
Brandon Ong
Brandon Ong@bytedunks·
Congrats on the Win! Creative way of probing model internals in layman terms.
ilham@ilhamfputra

we (w/ @Aufa_HR) won 1st place at the @GeminiApp Singapore Hackathon! your agent has taste. it's leaking into yours. Gemini's favorite color is indigo. Lyria's favorite genre is hip-hop. if you're not overriding these defaults, your agent is designing for you.

English
0
0
2
170
Brandon Ong
Brandon Ong@bytedunks·
China optimizes for number of robots sold. incentives drive outcomes👌🏻
Sourish Jasti@SourishJasti

1/ General-purpose robotics is the rare technological frontier where the US / China started at roughly the same time and there's no clear winner yet. To better understand the landscape, @zoeytang_1007, @intelchentwo, @vishnuman0 and I spent the last ~8 weeks creating a deep dive on humanoid robotics hardware and flew to China to see the supply chain firsthand. Here's everything we've created + our takeaways about the components, humanoid comparisons, supply chains, and geopolitics👇

English
0
0
1
174
Brandon Ong
Brandon Ong@bytedunks·
To understand what it takes to build a humanoid robot with model-based control, we finetuned @physical_int 's (PI) Pi05 model for our custom use case and environment. We incurred ~$10K in hardware costs, compared to the typical ~$20K set up (DROID/ALOHA). Here are the lessons and challenges we faced building the first working prototype (shown in the video) in 3 months. Part 1: Hardware, Software, Model Selection, Custom Embodiment, Inference, Embedded Hardware, Hierarchical Planner Part 2: Model Evaluation, Data Collection, Model Training, Simulation and Teleoperation We hope sharing our experience accelerates the learning of others who are in a similar starting point.
Brandon Ong@bytedunks

x.com/i/article/2018…

English
3
11
110
16.2K
Brandon Ong
Brandon Ong@bytedunks·
@Amank1412 @dabit3 Would be interesting to search this across company descriptions! Something like “AI agent for enterprise context” and search similar companies who pursued a similar problem space
English
0
0
0
216
Aman
Aman@Amank1412·
Someone built Startups.RIP a directory of 5,700+ failed YC startups with post mortems, deep analysis, and rebuild plans so you can revive dead ideas and turn them into new projects.
English
176
1K
9.8K
858.4K
Brandon Ong
Brandon Ong@bytedunks·
be distinctive while trying to be right, and don't die (or go to jail)
English
0
0
0
98