Pinned
Trevor Xu
139 posts
Cofounder of Pancrest Capital, incubating @DowProtocol | Founder @TaggerAI | National Winner @UNSW ProgComp, @austmathstrust
Melbourne, Australia
Joined March 2022
- It’s been a very busy period putting the final touches to our project before its long-awaited launch. I also had a taste of the most “interesting” (traumatic) experience of manually labeling data for 13 hours straight to feed our pre-training system. To put this dramatically,
- When the Tea App accidentally left its Firebase bucket wide open, 72000 user photos - including thousands of ID selfies - were scraped and splashed across the open web. Three things caused the incident: unencrypted files, no access‑control list, and zero audit trail. Tagger’sThe drivers licenses leaked today from the tea app have been uploaded to a searchable map.... this may be the worst PII leak I've ever seen lol
- "Everyone wants to do the model work, not the data work." This sentiment mirrors the earlier days of spreadsheet analyses, where the excitement lay in training models rather than the meticulous process of feature engineering. Yet, the reality remains that the investment in data
- The Quiet Doors I Kept Knocking On (I’ve added a Chinese translation in the second half) When I started Tagger, most advice I heard was to chase headlines. @unicornverse_io told me to chase fundamentals. Ann @A_unicornverse kept reminding me: build real utility, let the work
- Google cuts ties with Scale AI after Meta takes 49% stake sparking major shake-up in AI labeling market: • Google shifting $200M annual labeling spend away from Scale. • Microsoft, xAI, OpenAI also backing off. • Competitors Labelbox & Handshake already seeing demand
- The current discourse around multimodal research, often emphasizes the development and refinement of models. This perspective, while invaluable, tends to overlook a foundational challenge: the inherent issues within the datasets these models are trained and tested upon. I argue
- Scaling human resources to obtain high-quality data for training LLMs presents a uniquely challenging aspect of AI development, markedly more complex than scaling machine resources. Unlike machines, human output varies significantly due to numerous factors, including motivation,
- Came across an article today on the development of advanced techniques for whole-cell segmentation in tissue images. doi.org/10.1038/s41587… As someone who is founding businesses in the AI and data industries with a brief background in biomedical engineering, let's take a
- For the past 15 years, much of the activity has been speculative, fueled by promises of rapid gains rather than genuine utility. Now, the goal is to transition toward real-world adoption, both by enterprises and consumers. But there are clear obstacles to overcome. We need more
- We are honored to be a part of the #USD1 ecosystem and as winners of the @WorldLiberty, @BUILDonBsc_AI, and @PancakeSwap USD1 competition. We appreciate the endorsement and support from @worldlibertyfi and @ZachWitkoff. This validates an essential component of financialCongratulations to the winners of @worldlibertyfi trading competition @EGLL_american @liberty_bsc @TaggerAI @LorenzoProtocol. A big thank you to our co-sponsor @BUILDonBsc_AI and supporting partners @four_meme_ @Aster_DEX @lista_dao @BNBCHAIN @PancakeSwap. Let’s keep building












