Henry Mao

I'm currently putting together a newsletter (vivianchan.substack.com) to share the latest startup and VC jobs in the Deeptech space. Quite a few people have asked me for intros or to tap into my network, so I thought I’d collate everything and share it with everyone. If you or your portfolio companies are hiring, let me know and I'll include it in the next edition! 🙌 Examples of available roles: AI Research Engineer, FutureHouse (AI, Non-profit research org backed by Eric Schmidt), San Francisco, US Head of Operations - AimHi Earth (Sustainability, Seed), Remote Clinical Laboratory Scientist - Juno Bio (Bioinformatics / Microbiomes, Seed), Oakland, California, US Deeptech/Climate Investor - Founders Factory (VC), London, UK or NY, US Bioinformatics Consultant - Deep Science Ventures (VC), London, UK VP Finance - Space Capital (VC), NY, US 📩 More in the first edition: vivianchan.substack.com #Deeptech #Startups #VentureCapital #JobOpportunities #Networking #TechJobs #Innovation #Entrepreneurship #Hiring #Jobs #CareerGrowth

23

5 Comments

I'm launching a FREE GPT-4o-powered healthcare chatbot with knowledge of clinics and hospitals in Japan on June 12. This is me forcing myself to put this project out there instead of personal use. It's working fine I did not want to release a chatbot version but something with another layer where the user doesn't need to ask questions but push some buttons. Now I want to remove it from my head and have it out there. What are my assumptions about the users? - could be immigrants in Japan - tourists Why use this instead of Google? - its knowledge is tailored for healthcare facilities in Japan - it is also multilingual This also may not be the best platform to announce this so if you know a channel/group that will benefit from this let me know. Cheers

69

12 Comments

What's next for AI agentic workflows? Andrew Ng founder of DeepLearning.AI and AI Fund, has a super interesting take on what's next for AI agentic workflows and their potential to significantly propel AI advancements—perhaps even surpassing the impact of the forthcoming generation of foundational models. What's in? - AI agents: Iterative, agentic workflows enhance performance. - Reflective agents: Self-assessment improves code quality iteratively. - Multi-agent collaboration: Diverse agents boost complex program generation. - Fast token generation: Rapid token output crucial for iterative workflows. - AGI journey: Agent workflows propel progress in AI development. View it here: https://lnkd.in/dYu6daYr. --- To explore more: I created a "How to with AI agents" series for decision-makers exploring AI agents’ capacity to revolutionize how we work, live, and interact with technology: https://lnkd.in/dkTZ5RVj. Where I ask questions like: - How do we perform better with AI agents? - How do we better target the underserved? - How do we democratize AI? - Does a higher level of automation in AI mean less human control? For more, follow Wild Intelligence.

3

Super insightful talk by Tengyu Ma on RAG existing techniques, what is coming in the field and what they are building at Voyage AI. Super happy to also see how he is backing up Zylon by PrivateGPT's approach for privacy when he has been asked about how to secure private data when using RAG: "On-Premise deployments are the only solution to achieve it and you can do it with pretty affordable infrastructures built on top of cheap GPUs". #AIAREALL

7

1 Comment

Loved this insight on Risk from the esteemed Terence Tao In times of great uncertainty, the relative advantage of "playing it safe" diminishes because no option can completely eliminate risk. Paradoxically, in such situations, taking bold actions can become more rational, as decisions should align with the level of external risk in the system. This highlights that uncertainty fundamentally reshapes the balance between safe and bold choices, requiring a more nuanced approach to decision-making. Many decisions in life involve the tradeoff between risk and reward. Perhaps one has to choose between a "low risk, low reward" course of action that plays it safe but does not achieve any big wins, or a "high risk, high reward" choice that could potentially give greater benefits but also greater downsides. Risk management is a complex task, but it can initially be explored with very simplified models, always keeping in mind George Box's dictum that "all models are wrong, but some are useful." For this simple model, one can assume that each course of action comes with two numbers: the "reward," representing the average positive net gain, and the "risk," interpreted here as the standard deviation from the mean. A "safe" action might have a reward of 5 and a risk of 3, written as 5±3, while a "bold" action might have a reward of 9 and a risk of 10, written as 9±10. Given the general preference for risk aversion, one strategy is to minimize "value at risk"—the value one might lose in a bad scenario. For the safe action (5±3), the value at risk is 3-5 = -2, indicating that even in bad luck, a small gain is likely. For the bold action (9±10), the value at risk is 10-9 = 1, suggesting possible losses. This favors the safe action. When external shocks are introduced, the analysis changes. Suppose a fixed shock of -5 uniformly affects outcomes. While this lowers rewards, it does not alter relative decision-making. However, with uncertain shocks (e.g., -5±10), risk is recalculated using the law of additivity of variance. For uncorrelated risks, the safe action moves from 5±3 to 0±√(10²+3²)=0±10.44, while the bold action moves from 9±10 to 4±√(10²+10²)=4±14.14. This reduces the relative value of "playing it safe." In times of great uncertainty, no option remains truly safe. Paradoxically, bold actions can become more rational as personal risk tolerance aligns with systemic risk. However, this conclusion depends on the assumption that external shocks add to risks rather than multiply them. If risks are multiplied (e.g., converting 5±3 to 5±6), the safer option becomes preferable again. Behavioral economics highlights how human decisions deviate from models, as external shocks psychologically distort decisions. Nevertheless, there is a rational basis for altering choices when shocks themselves carry risk. The broader insight remains: uncertainty reshapes the value of safe versus bold decisions.

25

1 Comment

This - AI companies are having to build the 'brain' AND how that brain 'interacts' with a world built only for humans in mind. CloudCruise is abstracting away second bit so we can focus on nailing the first bit.

26

1 Comment

It is easy for us at Pints.ai to say 1.5 Pints is just amazing and gives a new framework to go about pre-training LLMs. So I asked the Goliath - GPT4o - to tell me what really sets 1.5 Pints apart. Keen to know how we trained a model that beats Apple and Microsofts' SLMs, in just 9 days (and definitely for much less $$$ than what they spent on a PR release for their models) - drop us a note. Jerome W. , Calvin Tan - if anyone has any Qs __ what is unique about this? anything that truly that stands out? The paper you provided, titled *"1.5-Pints Technical Report: Pretraining in Days, Not Months – Your Language Model Thrives on Quality Data"*, stands out for several key reasons: 1. Compute Efficiency: The "1.5-Pints" model is trained in just 9 days, which is a significant reduction compared to the months typically required for training large language models (LLMs). This is achieved without sacrificing performance, positioning it as a faster, more efficient alternative. 2. Data Quality Focus: Unlike many models that prioritize large datasets, 1.5-Pints focuses on data quality over quantity. Its pre-training dataset consists of 57 billion tokens of carefully curated content, such as research papers, copyright-free books, and parliamentary debates. This results in a model that outperforms peers trained on significantly larger datasets. 3. MT-Bench Outperformance: Based on the MT-Bench benchmark, which mimics human judgment, the 1.5-Pints model outperforms notable models like Apple’s OpenELM and Microsoft’s Phi despite being trained on 15-25 times less data. This highlights the success of its data-centric approach. 4. Reduced Environmental Impact: By shortening the training period and focusing on quality, the model also aims to reduce the carbon footprint associated with training large models, a growing concern in the field of AI. 5. Innovative Use of Mistral Tokenizer: The authors adopted a modified Mistral tokenizer, which is shown to outperform the popular Llama-2 tokenizer in terms of compression efficiency (3.61% fewer tokens), improving the model’s efficiency and performance. 6. Llama-2 Architecture Modifications: Although based on Llama-2, the model includes several key architectural modifications, such as an enlarged Multi-layer Perceptron (MLP) and Grouped Query Attention to further boost efficiency and speed. 7. Open-Source Focus: The paper’s results, code, and datasets are open-sourced, providing transparency and encouraging the research community to build upon their work. This open-source approach ensures reproducibility and contributes to advancements in the field. In summary, the paper’s emphasis on speed, data quality, efficiency, and reduced resource consumption are the key factors that make it unique. Its ability to outperform models trained on much larger datasets through a more efficient and targeted approach is a notable contribution to the field of language models.

41

1 Comment

Underappreciated milestones in AI - Pt. 5 𝗧𝗵𝗲 𝗙𝗶𝗿𝘀𝘁 𝗔𝗜 𝗪𝗶𝗻𝘁𝗲𝗿 (1973-1980) Not long after the Perceptron Mark 1 (Which we covered earlier in this series 😉) Artificial Intelligence began under-delivering. Particularly within machine translation. Leaving a sour taste around AI. Moderate developments like ELIZA, Helped restore some faith, however, soon Martin Minsky began criticizing the OG, Perceptron Mark 1. In a book known as Perceptrons published in 1969, Paired with the Lighthill Report in 1973. Spelling doom for AI funding. Leading to: 𝗔𝗜 𝗪𝗶𝗻𝘁𝗲𝗿 Check out the Carousel Below! Understand what AI WInter means Why it was special, And! The groundwork it laid! Also, it didn't last forever, in 1980, there was a boom 😮 - This will be my last AI in history post for a little bit. Unless you all want to see more! ⬇ Let me know in the comments Make sure to: 👍 Like ♻ Repost 🔔 Follow me for more Thank you for reading! #AIHistory #Perceptron #AI

36

40 Comments

Burnt out? Lost your tech mojo? Don't panic. Here's how to transform from code zombie to passionate innovator: 1. Embrace "productive laziness" • Cut back hours • Focus on impact, not input 2. Become a tech tourist • Explore new domains • Attend tech meetups 3. Break stuff (strategically) • Start a "destroy and rebuild" project • Gain insights by reverse-engineering 4. Teach a tech-illiterate friend • Explain complex concepts simply • Rediscover your "why" 5. Go on a "tech detox" • Disconnect completely for a week • Return with fresh perspective Rest isn't for the weary, it's for the wise. Which of these will you try this weeks Commit in the comments! 👇 *** P.S. Tomorrow, I'll reveal how to amplify your reignited passion. You'll kick yourself if you miss it, so hit the 🔔 on the way out.

25

15 Comments

There are many 🫣 hidden Swift functions that can still be accessed in Apple frameworks like SwiftUI. Finding & using these functions is actually straightforward once you know how 🕵️‍♂️ In the latest Emerge Tools post Noah walks through: - How Apple uses SPI to hide functions - What tools you can use to find them - How you can use these hidden functions (just like we do) https://lnkd.in/gvzqD8MK

65

4 Comments

Google has I/O, Apple has WWDC, and now OpenAI has DevDays. What's different this time? OpenAI has made a significant move by bringing its core developers from San Francisco to Asia. The event was the perfect size. There were enough people to create serendipitous encounters, yet ample space to relax (or handle those emergency calls...). The pace was just right, balancing 30-minute sessions with 15-minute breaks, which kept the energy high without being overwhelming. Interestingly, despite ai.com now redirecting to ChatGPT.com, the demos focused heavily on API integrations rather than ChatGPT itself. Here's what stood out: - Efficient Software Creation: Developing software with minimal supervision has never been more streamlined. - Cost-Effective Testing: Enhanced methods to test results are significantly reducing costs with evaluation and distillation. - Advanced Conversational APIs: Voice commands are evolving beyond personal assistants to building complex processes through conversation. The key takeaway? These advancements empower individuals to achieve what once required entire teams...quite a paradigm shift!

45

Southeast Asia's Uber-esque superapp, Grab, has developed a tool that allows its employees to build large language model (LLM) apps without coding. The company this week described the tool, which it calls Spellvault, as empowering "everyone in the company, even those without coding skills," to "create LLM apps that have access to the rich repository of Grab data." The superapp claimed custom apps can be built in as little as five minutes. It gave two examples of groups that had developed tools. Its insurance division built a chatbot to provide information on policies, inquiries, and claims procedures, and its localization group developed a translation app. Spellvault operates on the front end through a three-step process. Users start by defining the model's task, such as assessing product initiatives for potential policy risks. Then they upload relevant data into the Knowledge Vault, a repository where Spellvault stores structured information. Finally, users can specify APIs if the model needs additional data from external sources. On the back end, Spellvault uses retrieval-augmented generation (RAG), according to a post on Grab's tech blog last month. The AI framework integrates a retrieval step before generating responses, which is claimed to improve accuracy and relevance of LLMs... #artificialintelligence #technology #employee #ai

41

19 Comments

MOST Gen AI POCs will fail to make it to production. Why? 1. We have had some potential clients expecting us to build POCs for free ("You can put our logo on your slides, but we won't pay") - frankly if you're not willing to pay for a product, you haven't even given it enough thought or consideration. It's just wasting everyone's time. And Gen AI is not a cheap solution, especially when it's custom build. 2. There is no clear understanding of 'cost of compute'. "Let's just do a POC now, and we will figure deployment later". There's a reason why NVIDIA is a 3 Trillion $ co - there's a chip demand-supply mismatch and cost of a product (compute) in a a supply restricted market is going to stay elevated in short run. No clear deployment plan post POC is an another brilliant dead end. 3. No Data strategy - you just can't use ALL of your data on one single LLM. One of the biggest problems we have seen (and are addressing for our clients) is using PII data in Gen AI workflows. People we meet in Insurance, Banking, family office etc tell us there is more data they can't put onto Gen AI workflows, than the one they use. Perfunctory data strategy, or using simulated data to push through a POC won't get you anywhere meaningful. 4. One LLM to rule over all mistake - Way too often we come across "Lets just use the most powerful LLM and do all of this". When the fact is, some LLMs are better for some use cases you are trying to solve. Keep aside the compute cost, deployment flexibility, ownership of data etc. LLMs are tools - just like hammers and screw drivers. You can, if you wish to, hammer a nail with a screw driver, but… 5. Too many POCs - how many POCs are too many? We met someone who showed us a deck of 58 POC ideas that their company is working through (I am understating that number). I was left wondering, isn't it better to just get team leaders from different functions into a room, have a simple discussion with a whiteboard and list what problems would they like to eliminate? 58+ POCs is just lazy - the list was built by prompting chatgpt - "give me 5 dozen Gen AI POC ideas for an Bank" Gen AI is not magic. It won't make all your problems disappear over a weekend, it won't suddenly make your sales guys super heroes, or get your costs down by 50% by the time you go on your Christmas holidays. It is a super powerful tool for sure, and instead of wasting time and money on 5 dozen POCs, it might just be better to 3-5 solid ones with a clarity of how to execute till the end.

19

2 Comments

You may remember the scene in the movie Avengers: Infinity War in which Doctor Strange looked into the future millions of times to find every possible scenario of an alternative reality. Unfortunately, time travel is not yet possible in real life, but in the near future, you could use AI agents to perform complex reasoning tasks to simulate complex business scenarios. Imagine you are a small law firm representing a client in a complicated case involving multiple parties (government regulators, corporations, local authorities, etc.). You could create various AI agents representing each party and their incentives to play out various scenarios to determine the best action plan for your client. This would enable the wider public to access top expertise. Or imagine a hedge fund that would use AI agents to determine how various market events will play out by simulating the actions of various market participants. You already have the tools for multi-agent workflows at your disposal. For instance, LangChain offers a framework to build and run dynamic, interactive multiagent simulations quickly. Admittedly, challenges persist, such as the reliability of multi-agent collaboration, but the field is evolving rapidly. If you are interested, you may go through the short video presentation below by Andrew Ng, where he provides an overview of agentic AI and what is next for AI agentic workflows - https://lnkd.in/gxyf_DXg

8

Pumped to see Airbyte 1.0 release! Tons of awesome announcements, but some of my favorites include: - A new AI Assistant to build connectors in one click from API docs links - A Connector Marketplace where you can contribute or edit connectors in one click again - Support for many GenAI workflows

8

Real or Deepfake? You tell me... 👇 AI imagery is getting so good, it's hard to tell what's real. Genuinely, I don't think people realise that yet. And it's becoming somewhat of a problem. According to an article from the Economic Times... A Hong Kong MNC suffered a $25.6m loss due to a deepfake scam. The perpetrators utilised AI to mimic company officials, orchestrating convincing video conference deception. So you tell me, which of these visuals are real, and which are fake? Good on you if you get them all right. P.S. Let me know in the comments how many you got right ⬇️ 🟩 🟩 🟩 🟩 🟩 🟩 🟩 🟩 If you want to know the dos and don'ts around AI... Register for my FREE AI webinar that's happening TODAY: https://lnkd.in/eV6E7yTv ♻️ Don't forget to repost and follow me Tom Head as well.

121

118 Comments