gpt-oss + ultravox, should be an incredible combo for oss voice ai!

We're hosting two Model Context Protocol themed lightning talk nights in March, save the date! March 18 in SF with @CloudflareDev March 27 in NYC with @SignalFire

New month, new @SignalFire AI Lab event! Please join us as we welcome the CEO of @Horizon3ai Snehal Antani, CEO of @robusthq Yaron Singer, and Head of AI of @PaloAltoNtwks Nandan Thor. https://t.co/0hRU8vC2je

4. @AnthropicAI also unable to get me live data, and a bit oddly asks me for the answer rather than giving me the information I want 5. @MistralAI also unable to get the right answer and a bit slower 6. @xai with no free offering is an immediate no

Final results: 1. @GoogleAI the clear leader with the fastest answer immediately 2. @OpenAI gets us to the right answer but definitely slower and requires follow up 3. @AIatMeta is blazingly fast and the least verbose, but no access to live data

(11) Now to try out @xai with Grok 2. No free option so need to pay $8 a month, no thank you

(10) Another 8 seconds later, we're unable to run a live search and so get stuck. A bit odd that it hung for a while before responding

(9) Moving onto @MistralAI's Le Chat, we default to their Large 2 model which has a cutoff all the way back to October 2023 with the wrong answer in 4.5 seconds

(8) Trying with the same follow-up prompt, we are told that Meta's hosted Llama model does not have access to live data and aren't shared a cutoff date in 3.5 seconds. Still super fast but not very helpful

(7) Up next is @AIatMeta right after today's release of their latest 3.2 Llama models. We change the default to the larger 405B parameter version, and get the same wrong info but the fastest so far at just 2 seconds

(6) Onto @GoogleAI, we are offered a free trial for Google One which grants us access to Gemini Advanced's 1.5 Pro model. The same request does an immediate live search, and gives us the right answer in just 3 seconds. Very impressive!

(5) When we try to re-prompt, we are told that the last update was performed in April 2024 but unable to run a real-time search after 5.5 seconds. Oh well

(4) Now onto @AnthropicAI, we send the same request to Claude 3.5 Sonnet. A bit quicker than OpenAI at only 4 seconds, but still get the wrong answer

(3) After re-prompting, it takes 9 seconds to run a live internet search but we end up with the right answer this time. Not the best experience but we get there

(2) Starting with @OpenAI, we try out @ChatGPTapp which defaults us to their GPT 4-o model. 5 seconds later and we are told @miramurati is still the CTO

(1) Thought I'd use today's announcement of @miramurati leaving @OpenAI to run my own evals

Please join the @SignalFire team as we discuss all things data quality/curation with @AlexGDimakis and @arimorcos on August 21st! Register here:

Startup founders building LLM solutions need to understand the needs of established companies working in regulated industries. Otherwise, they will not be able to effectively partner with such firms to meet their evolving needs. Unusual investor @RyanMWexler recently chatted

Today dbt announced column level lineage... but it's not open source and only available for paying cloud users. dbt is using SQLGlot (an open source SQL framework that my company @TobikoData maintains) to power this newly gated feature. SQLGlot's lineage wouldn't exist without

Happy Friday. Our secret is that https://t.co/IEvUZCoxb8 GPUs are now available for everyone to try, but you didn't hear it from me. https://t.co/lEIIRTkoqo