Filip Vítek

I was impressed by Lllama-3's performance at 8 billion parameters. Well, Microsoft just released a model half its size with comparable performance. The model is Phi-3, and the smallest version is 3.8 billion parameters, yet achieves comparable performance to GPT-3.5 across many (but admittedly not all) benchmarks. The Phi series of models use highly curated data, including a lot of synthetic data generated by GPT-4, to achieve these great results. What this means for the world: Powerful AI that can run locally on devices including phones, and be very tailored via retrieval-augmented generation or fine-tuning to individuals, companies, and tasks. I have to update my understanding of the world to comprehend how you can compress an internet's worth of information into 3.8 billion parameters. Paper is here: https://lnkd.in/gYNswkcs

2

7 Kommentare

Important: OpenAI is (I think fairly) excluding o1's reasoning from its API output, which has significant implications for synthetic data generation and generative AI market dynamics. Right now, people use APIs from OpenAI and other companies to generate training data for smaller, faster, cheaper specialized models. For example, if someone wanted to train a model to extract contact information from a website into JSON, they could have GPT-4o generate 10,000 examples of doing this, then use that to train a small model that they owned. This can be useful for situations such as when a big, general-purpose model is overkill for a very focused use case. Or, perhaps because you want to run a model in a pipeline and not rely on a third-party vendor's uptime. However, it also contributes to rapid commoditization, because people can also use such output to train or fine-tune competing models, even though terms of use often prohibit this. For example, if you wanted an open-source model to be as good as GPT-4o at an economically useful task, you could fine tune it on GPT-4o output for that task. With o1, OpenAI is changing this. While the o1 API will output the model's final answer, it won't output all of the reasoning that went into that answer. This will make it impossible to mirror o1's ability to reason without creating a similar training process to what OpenAI uses, which as I understand it is a combination of chain-of-thought reasoning, reinforcement learning, and training on resulting synthetic data. This could have a SIGNIFICANT impact on market dynamics. Currently, GPT-4-level capability is a commodity. Many models are at a GPT-4 level, and companies are battling on things like price and speed. But I don't think this will be as much the case for o1-level reasoning models. We may see models at this level from the biggest labs—OpenAI, Google, Anthropic, xAI, Meta—over time. But I think we'll see way fewer companies able to compete and a slower pace to the level of commoditization we now have at the top model level.

15

5 Kommentare

For everyone getting hot under the collar about the new "computer use" beta feature in Anthropic's Claude 3.5 Sonnet model (currently scoring between 14.9% and 22% on the OSWorld benchmark, which attempts to evaluate models' ability to use computers like people do)... You do know about RPA, right? 😉 Snark aside: These new features are really cool research directions that provide tantalising glimpses of what might be possible in the future. But if you want to hook up a LLM to take actions in apps today, do it using tools that are proven to actually work and scale. Do it via APIs, where possible/practical, or via RPA.

34

5 Kommentare

WHAT YOU NEED FOR CAUSAL ANALYTICS IN GTM It's interesting to watch how the market moves. Last week, as though everyone was having the same thoughts, we got lots of questions about what's necessary to get going with causal analytics. The nut of the questions, of course, is this: "hey, #CausalAnalytics sounds like the answer we need, but is it going to be too much for my team to handle?" Here's what you need to be thinking about: #MindsetChange. The core logic is simple but profound: GTM is a dynamic, relational, and probabilistic set of relationships that drive Value. Our collective goal is to make better and better bets to drive more and more value. Marketing does this by making Sales more effective and efficient than Sales could be without Marketing. This is the #MarketingMultiplier. A norm in B2B is that Marketing makes Sales 8X more effective and 5X more efficient. Marketing brings a tremendous leverage to the entire customer journey, enabling Sales to do #MoreDeals, #BiggerDeals, and #FasterDeals than it could by itself. Historically, B2B GTM teams have thought about their activities and their value in linear, orchestrated, synchronous ways, focused on near-term effects like MQLs. This not only is not how it works, but means that 90% of the value Marketing creates for Sales goes unrecognized. Marketing and Sales create collective value in nonlinear, synergistic, and asynchronous relationship, both with each other AND WITH CUSTOMERS. #Data. The big thing to say here is that no one has perfect data, so it's important to nuke that idea. The data you will need is related to the questions you want to model. A classic example would be "we want to understand the effects of all of our GTM investments on Revenue generation, net of time lag and relevant external factors." This means you'd need to muster time series data on the performance of most, if not all, of your investments, as well as relevant marketplace headwinds or tailwinds that you believe might increase or interfere with the effects of those investments. #Modeling. Most GTM teams are not going to do this work themselves unless they have a command of basic data analytics and logic. The key today is that causal analytics have moved away from Waterfall, mega-model approaches to Agile, minimum viable models that can be built in hours not days or weeks. This approach enables automated recalculation of models at whatever cadence is required, giving you a "GPS" understanding of your GTM performance. #People. A normal deployment of ProofAnalytics.ai requires ~6 hours of data analyst time per week, based on 20-30 models. Once the models are running, the analyst monitors and maintains them, much as a DevOps pro would manage software or systems. Analysts and business teams can collaboratively create models in Proof with exceptional speed and alignment in order to learn and validate hypotheses across time. Normal time to value is less than 30 days from the start of implementation.

19

12 Kommentare

Breaking Language Barriers: The GPT40 App vs. Moshi 🚀 Just tested the real-time translation capabilities of The GPT40 app, and WOW, it's a game-changer! 🌍✨ 🔍 Compared to Moshi, which I tested recently, here's the scoop: Response Time: Moshi is quicker, but... Accuracy: GPT40 wins hands down. It's like having a human translator in your pocket! 🧠📱 Bidirectional Translation: GPT40 doesn't just translate; it does so BOTH ways seamlessly. Perfect for real-world conversations with non-English speakers! 🗣️↔️🗣️ Imagine traveling or working internationally and being able to communicate effortlessly. No more awkward misunderstandings or lost-in-translation moments. 💬🔄 If you're in need of reliable translation, whether for business or personal use, give The GPT40 app a try. It’s transforming the way we connect globally! #Innovation #Tech #Translation #GlobalCommunication #GPT40 #LanguageBarriers #RealTimeTranslation #AI #TechTrends #BusinessTravel #InternationalRelations #DigitalTransformation

2

2 Kommentare