SpeakEZ.ai

We sure would enjoy this stage and lighting for being founded on human centered design since 2021.

?? ? ?? May the wonders never cease Teams finally has #fsharp markup. #dotnet

?? We're glad to see this #machinelearning project get renewed attention. It not only has an API similar to #PyTorch but it blends in the additional power and precision of #fsharp for making quick, repeatable work for a variety of #datascienct tasks. https://lnkd.in/enk5zwQ2

?? I'm helping to revive an #AI and #machinelearning project in the #fsharp ecosystem we've renamed "Furnace". It has some incredible foundations in the original work and has been 'idling' for a while. As such, there's updates to the underpinnings needed, but once we get things fired up ?? (pun!) ?? there will new releases and several optimizations in the works. Stay tuned!

It tracks that most VCs spout senseless word salad and rely on dart-board law of averages to occasionally hit a bulls-eye ?? so replacing them with a statistical model would likely be an improvement as at least there would be less ego involved. https://lnkd.in/ed3eFncr

?? I understand why these kinds of pseudo-FUD posts get traction, but I don't believe it's wise to play this game over the long haul. For a bit over a year SpeakEZ.ai has had domain-tuned generative models parsing various structured data sets - including standard SQL *and* cypher queries. There's nothing noteworthy about what Uber or anyone else is doing other than *everyone* should be using similar techniques if they have a reliable corpus of data over which to build this kind of tool. ?? A constellation of retrieval methods will be part-and-parcel of any domain-based solution that wishes to maintain sovereignty over their own data. If we're going to get the most out of AI we need to have as much clarity and transparency into the mechanism as possible, and handing it over to someone's "GPT" tool is not part of a sustainable future. (it would be great for *them* but not so much for anyone else) ?? The attention value of these kinds of posts will collapse to 0 faster than the value of aspiring hyperscale AI shops. ?? ?? 0?? Let's get back to the real work - as there's so much yet to do!

In case you're wondering if #cloud based #hyperscale #AI is just another bubble - we submit for your consideration - a startup is trying to spend $5 million on an F1 race team sponsorship instead of hiring staff or building infrastructure. ?? ?? https://lnkd.in/e8GBnhZa

One reason we built this proof is to show that #analytics fundamentals still apply to grounding and aligning #LLMs suited to your data. Foundations such as cataloging, indexing, surrogate keys and business transforms, they apply now more than ever. In future videos we'll "peel the onion" to show how we made this public data useful for high speed inference and high leverage tasks that reveal insights - in medications, in prescribing patterns and in the prescribers themselves. https://lnkd.in/esugsjH7

We're glad to see DeepSeek AI publicly establish the role of reinforcement learning in deep model training. We've been on that track since our founding, and quietly working on it ever since. In light of our own work it's been interesting to observe "frequentist" shops like OpenAI shuffle sideways into using RL without "giving away" that their original (and still primary) thesis of "more data is always better" is inherently, tragically flawed. Our process has been in "stealth mode" for a while as we're going even further than DeepSeek - with the ability to record knowledge accumulation around models that eventually yields follow-on refinement. It's still a subject of research, but the gist is to find the inflection point where the knowledge base growth achieves a critical mass such that "baking it in" (which DeepSeek refers to as "cold start") benefits the CoT processing internal to the model. This human-to-machine-and-back cycle is part of our "Systm2" platform, and as our own roadmap develops we'll continue to welcome the rest of the #AI industry to the RL arena.

Ironing out an #fsharp based #AI feature for industry selection. There are several methods to select an Industry/Subindustry. "Smart Search" inference is fast - nearly as fast as the filter method - and is very flexible. Note when going from "Police" to "Police Benevolent Associaton" (sic) the inference updates the context to pick up trade associations, even through misspelling. ??