Alberto Savoia

Vishal Sachdev highlights the strategic integration of open source and proprietary tech in architecting tech stacks, developer ecosystems and resulting business models. The world class companies effectively balance value commoditization in open source and value capture in proprietary tech. This is the strategic challenge for companies operating in #blockchain and #AI. I believe those who operate at the intersection of blockchain *and* AI stand to win this strategic battle...

3

1 Comment

Enterprise Automations, RPAs, AI, Agents - where and how do they intersect and overlap? Great coverage in this article from Insight Partners on this topic: https://lnkd.in/ghiNGtC8

59

3 Comments

People keep getting this wrong. It is NOT about Deep Tech or Hardware. It is about System of Records, Platforms, IP and most importantly MOATS OVER TIME. Businesses that accrue moats over time have specific signatures and exist across verticals -- deep tech, biotech, health, saas. Thats our thesis 2048 Ventures

37

2 Comments

I had a nice catch-up with an investor this week who spoke about how AI start-ups have evolved their strategies from pure infrastructure plays to more verticalized applications. The rationale was simple: if there is no market for applications, who is buying the infra? Of course applications require more focus and prioritization on a smaller set of customers with more niche use cases. Many founders are resistant to that, because it lowers the TAM (in the short-term anyway) and shifts the focus from the technology itself to customer pain points and value props. But without a customer, there is no viable product, and without a product, investors are only willing invest for so long. Even Microsoft is not immune to this reality. No matter how much any company invests in a solution, they still need a market to pay for it.

2

1 Comment

Multi-agent systems have been steadily gaining adoption, but it’s important to understand that they aren’t a blanket solution.  In fact, they introduce a range of complexities around coordination, consistency, and coherence that single-agent setups avoid. In my conversation with Chi Wang, the creator of AutoGen, he underscored the importance of matching your product architecture to the problem you’re trying to solve. If your use case involves narrowly defined tasks, a lone agent may be the simpler, more efficient choice. Other factors to consider before making the single agent versus multi agent decision? Response speed, decision-making frequency, inter-agent communication needs, latency, and bandwidth. Check out my full conversation with Chi here:

39

9 Comments

Dystopia may be on the horizon… LLMs are only as good as the data on which they’re trained. That’s it, they can’t create anything new, but they can bring a wealth of knowledge and expertise distilled down into a simple interface for us to use. Companies like Meta and OpenAI and Google are training their LLMs on whatever information they have available to them. And this can, and already has, had disastrous effects. From racist chatbots trained on Facebook, to outright trolls trained on Twitter and reddit, without appropriate curation of training data, the results are staggeringly bad. So when it comes to regimes like China (and others) training LLMs based on their propaganda machines, the opportunity for abuse becomes outrageous. Leading me to conclude in this case that the wrong kind of politics will lead AI down an ever polarising direction. We’ve seen two AI chatbots talking to each other this week. Both 4o models, but I wonder what happens when those models differ and one is trained in an ideology not supported by the other. This doesn’t seem like a big deal until you imagine a world where your own personal AI out in the world is interacting not only with the websites you want to visit, but also other agents. When you do that, you’ll want your agent to not be manipulated by adversarial agents. Trained in how to spot a fake, ready to fend off the false claims one might make.. But unfortunately that might not always be the case, after all you could have some innocent person asking “how to make my cheese stick to the pizza better” with the AI telling them to “use non-toxic glue”. I know you’re not silly enough to fall for it, but I guarantee someone will. And if instead someone is asking a politically charged question, one which has a variety of viewpoints, don’t you want to have a balanced answer instead of whatever answer your government wants you to see? Yes you do.

6

2 Comments

Worth reading: Generative AI is not going to build your engineering team for you. So, so good - peppered with gem quotes and observations- chuckle. https://lnkd.in/gZBKtU4X Thanks to Eric Brooke for sharing.

16

Is Fine-Tuning an #LLM a waste of time? Perhaps the answer is: YES! 😁 Important info for practitioners, product leaders and investors. At the Tel Aviv University #Aiweek conference, Jonathan Herzig from Google showed some great results demonstrating that fine-tuning a large language model isn't quite what you think. It tends to cause the LLM to hallucinate more. This is important, since for over a year, I've been told by "experts" and investors (!?!) that all you need is to capture a unique dataset, fine-tune, and voila! A better #AI system! However, that flies in the face of reason on how these LLMs work, specifically, #transformer based #neuralnetworks. An LLM's architecture and mechanism is different from a deep-CNN-like model (like those used in #computervision), where internal layers have meaning, and therefore can be adjusted. A transformer (rather, the Attention mechanism) works differently and is more statistical in nature. Point is - the result makes sense, and it has taken over a year for people to start accepting this fact. it's a shame when you think about how much money has been wasted, silly #startups invested in, and how much data was "liberally obtained" (!), just to create wasteful hallucinatory models... Silly side note: please don't use light green on white in presentations... and use the Fullscreen mode 🤣

23

2 Comments

America may run out of power with AI's growing energy appetite 🔌🏭 The bottleneck for AI progress will shift from NVIDIA/TSMC's chip production to energy companies' capacity to power AI data centers. For example: ⚡️ Amazon's 1GW data center next to a nuclear plant https://lnkd.in/g4zqTwhG ⚡️ Microsoft & OpenAI planning $100B, multi-gigawatt "Stargate" AI supercluster https://lnkd.in/gWvnhZms To feed the AI beast, we'll need 💰💰💰 of investments in new power plants, massive data centers, and chip fabs that can churn millions of GPUs/year. Leopold Aschenbrenner's back-of-the-envelope analysis highlights the staggering power needs (https://lnkd.in/gfBc5xhb) 💡 Rumored $100B cluster = 10GW = small US State's power draw The intersection of AI and energy markets is 🔥🔥🔥 Think beyond chips and models. What's the NVIDIA or TSMC of the energy space? Nuclear? Solar? Exotic energy source? #ai #ml #llm #largelanguagemodels #gpu #cloud #datacenter #aicompute #energy #datacenterenergy #nuclear #greenenergy #startups #venturecapital

34

24 Comments

I fully support Andrew Ng's perspective. SB-1047 feels like blaming knife makers for kitchen accidents and crimes. We should focus on monitoring those who might misuse technology, not restricting the developers who create it. --- (b) Before using a covered model or covered model derivative, or making a covered model or covered model derivative available for commercial or public use, the developer of a covered model shall do all of the following: (1) Assess whether the covered model is reasonably capable of causing or enabling a critical harm. (2) Implement reasonable safeguards to prevent the covered model and covered model derivatives from causing or enabling a critical harm. (3) Ensure, to the extent reasonably possible, that the covered model’s actions and the actions of covered model derivatives, as well as critical harms resulting from their actions, can be accurately and reliably attributed to them. --- The internet might never have been invented if such laws were in place at the time. I hope the current version of the law is just a first iteration and will be improved by the attention and feedback of the tech community. #AI #Law

5

The Rabbit R1, the Humane, all these AI wearables seem like great products, for the wrong initial market. Market is full of 28 year-old boomers too attached to their existing tech. I truly believe when we look at the next generation of device we carry -- we shouldn't be appealing to techies and hackers. We should be looking towards the generation that is growing up with these devices. In reality, something like the R1 is literally perfect for a the curious 4-9 year olds of today who want to ask things like "what flower is this" or "how far is the sun". No grown ass adult is gonna point a camera towards a Cybertruck and ask “Yo what car is this”. One day that will be the case. But today, these products will need to be 100X better to ever trump over a smartphone. This will take time a lot of money and iteration. But, it will happen. Though, these products coming out today that are being dumped on...they're literally already perfect right now -- for kids. Children are naturally curious as hell. Parents want to manage screen time. A product that lets kids 24/7 ask the questions they're curious about seems... perfect. I started an online elementary school that had over 100,000 kids aged 4-9. Parents were always on the lookout for products that lets explore/learn without being attached to a screen. Market is there. Product is there. And, I think building for the new-gen (kids) will lead to more innovative products that can appeal to the masses in the long-run vs trying to appeal in the short run to the guy too attached to his Apple iPhone.

141

25 Comments

As academic based LLM benchmarks are losing their relevance -- there are 2 types of Benchmarks that will matter more in the B2B space: 👉 Crowdsourced user votes to rank LLM output (like Chatbot Arena) OR where the test DATASET is kept secret by the evaluator (e.g GSM1k from scale_AI) 👉  Enterprise use case specific LLM benchmarks (e.g produced by CRM, ITSM, HR SaaS companies) evaluated on real data. Because you can not game real customer use cases. #llm #benchmarks #llmeval #AI

11

1 Comment

A Google engineer saw the future in 2018: You may have seen the recently released email from Mark Zuckerberg to Peter Thiel go viral. It was popularized by the TechEmails dot com blog that surfaces emails that enter the public domain via big tech trials. While that was a fascinating email, I'm surprised that an earlier email did not go as viral. Eric Lehman, a Google engineer, foresaw the challenge that LLMs would pose for Google back in 2018. He warned the company to get ahead of it: "[BERT] had outperformed everything that we'd done — you know, dozens of engineers over roughly a decade. And so I sent this email to reflect on what the implications of that were for the future. It was quite a shock. I always felt that Google's — you know, its really big advantage is that by focusing on search for essentially so many years, we developed this deep body of theory and practices and we made a lot of mistakes but we learned over time. And I thought that was kind of our treasure chest. And then along comes this system and just like, no, never mind, beats everything. And at this point we had just seen that this kind of, sort of junior problem, web answers, had experienced this just dramatic disruption, where machine learning made everything previous kind of irrelevant. And the thought I'm considering here is that that's going to happen to web search, too. I was guessing that these kinds of advances would — yeah, just kind of clear the table of all past work." The blog also quoted Aravind Srinivas, CEO of Perplexity "In the first few weeks of launching Perplexity, an early OG Google employee told me the same thing as this: ‘don't worry too much about getting a lot of users to compete with Google. You're living in an age where unsupervised learning from raw internet text works. you do not need as much click stream data to build good index and ranking, and there lies your opportunity’" Google's market share doesn't appear to be hurting that much...yet... But it certainly poses a challenge.

5

1 Comment

Key pain points of LLM evaluation well summarized by Andrew Ng: - Lack of testing dataset / expensive to curate manually - Difficult to create good metrics to capture nuances - Expensive to run eval (can easily cost ~$200 / day in his estimate) - Time-consuming to run - makes iteration slower If you've been reading our blog posts, you would know how Relari / continuous-eval are addressing these pain points!

24

"..Although it’s always helpful to improve documentation for developers, many (myself included) prefer to dive in and learn while doing." -- an excerpt from the quoted post below. That quote from Jon Udell highlights a core challenge with documentation: determining the target audience. Unfortunately, trying to cater to everyone often leads to disenfranchising some, which can turn documentation from a tool for spreading goodwill into one that spreads ill-will. Large Language Model-driven chatbots offer a more adaptable solution to this age-old problem by providing a companion that aids working through the problem, courtesy of its natural language interaction modality. Jon and I have known each other for what feels like eons in computer time. Even when we haven't spoken directly, I feel we're often on the same wavelength about the evolution of software. For instance, I was addressing a similar issue with production documentation earlier today and turned to our #OPAL agent to expedite finding a solution, as illustrated by the following session transcripts. [1] https://lnkd.in/eWrrbXMG -- where the &t url parameter puts the page view in animated teletype mode (approx 120 per msec per typed word interval). [2] https://lnkd.in/eVZKE2vf -- conventional page view. #GenAI #AI #SmartAgent #UseCase

3

2 Comments