Bloom AI

Generic AI is harming student learning. One study from @Penn showed that it degraded student performance by 17%. So we built @BloomAITutor: AI that reinforces learning instead of bypassing it. If you want early access, reply and I'll DM you a code.

Link to press release: bloom.study/bloom-ai-unsw-…

We're incredibly proud to be working alongside leading researchers from @UNSWCOMPUTING, including Lorenzo Lee Solano, Jake Renzella, Sasha Vassar, Hammond Pearce and Andrew Taylor.

We're excited to announce that Bloom AI, in collaboration with @UNSW, has secured a $480,000 Ignite grant from Australia's Economic Accelerator (AEA) to develop "SPARK: Sovereign Pedagogical AI for Reinforcing Knowledge."

We need AI that works with students, not just for them. ✍️ Students can collaborate with Bloom through the Canvas Interface. Informed by feedback best practices and aligned with your institution's curriculum, Bloom nudges and guides, but the student is always in control.

Most AI tutors just give answers. Bloom AI Tutor does something different—it guides students to think critically.

Can AI tutors outperform traditional teaching methods? A preprint by researchers at @Harvard found that students using an AI tutor learned over twice as much in less time compared to those in an active-learning class. See my write-up here: garyliang.substack.com/p/ai-tutoring-…

Can generative AI harm learning? The answer is yes, if implemented blindly. A brilliant paper was recently published by @hamsabastani, @obastani, @Alpsungu, @HaosenGe, Özge Kabakcı and Rei Mariman, researchers from the @Penn and @BBIS_Budapest. Reading this study gave me even more conviction of the importance of what we're doing at @BloomAITutor. They undertook a randomised controlled trial on a group of high school mathematics students in Turkey. The students were split into 3 groups: - Control: No AI, just course textbooks and notes. - GPT Base: Access to GPT-4 chatbot similar to ChatGPT. - GPT Tutor: Access to a carefully crafted GPT-4 chatbot, with a prompt which contained detailed solutions and instructions to not reveal answers directly. Students' performance in the assisted practice problems substantially improved in the GPT Base group (by 48%) and in the GPT Tutor group (by 127%). The big result was that unassisted exam performance actually degraded by 17% for the GPT Base group. The negative effects were largely mitigated by the safeguards in the GPT Tutor group. My key takeaways: 1. Blindly implementing off-the-shelf AI tools may harm learning 2. We need AI specialised for education 3. Give students guidance on how to use generative AI 4. AI is like the calculator… but it's also not I also suggest some further avenues of exploration. Check out my write-up here: garyliang.substack.com/p/generative-a…

I started a Substack about AI in Education. My first article is about @Google's LearnLM and what we can learn from their fantastic technical paper, led by Irina Jurenka and Markus Kunesch from @GoogleDeepMind Check it out here: open.substack.com/pub/garyliang/…

I'm excited to be visiting San Diego to represent @BloomAITutor at the 2024 @asugsvsummit . If you're interested in how we use AI tutors in classrooms to improve student experience, please reach out!