thatsadude

Since Gemini 3.1 Pro, it's all benchmaxxed models. I recently tried Argon 4 and it failed to use tool correctly so much it's not worth bothering about Google's models anymore.

I have a completely opposite experience. DS4.1 and Mimo v2.6 take too long to finish a task, not worth it.

These investors are not that smart

All FF did was building a dataset.

https://arxiv.org/pdf/2402.03300 DeepseekMath was published many months before O1-preview.

DeepSeek invented the whole reasoning paradigm and keep pushing for innovation. I hope they get the success they deserve.

This is “google brain”

Funny to think about this when he got fired releasing a well-received product. On the other hand, Google released a subpar antigravity 2.0, full of bugs, almost unusable. The tech lead of Antigravity went on X to claim users don't know what they want.

So are my DSP/Audio ML career, I spent decades acquiring the expertise. Now at 40, I don't know what to do if I lost my current job. Very sad, but from economic POV, I also don't need junior engineers in my team anymore.

Nawh, they trained on test since Llama 2, no wonder.

On semantic VAD, I recommend https://github.com/pipecat-ai/smart-turn

I hope SpaceX succeeds

Nice read!

Has been a life changer for me.

See: https://arxiv.org/pdf/2411.03923

Same. I went to this thread and saw similar talking points brought up. Dangerous time for the US ahead.

Bad investment IMHO. Mistral was started by people who cheated on benchmarks with their Llama 1. It showed as they had the head start but fell far behind Gemini, DeepSeek and Qwen teams.

You can optimize prompt with MIPROv2 without examples (set the max number of examples to 0)

My go to framework. I wish we can use global metrics in DSPy, for examples, F1 score over the whole evaluation set (instead of a single query at the moment). The recent async support has been life saver.

I’m disappointed. Nobel price for neural network architects that failed the test of time..

Well more deserving than the author of Restricted Boltzmann machine.

True, that's a yellow flag for me.

I had a few problems with DSPy:

* Multi-hop reasoning rarely works with real data in my case. * Impossible to define advanced metrics over the whole dataset. * No async support

That plastic dish (parabolic reflector) is not cheap.

Recent ASR models are already robust to noise due to Spec augment and large-scale data. If you use these noise reduction services to remove noise, ASR models will have harder time to recognize denoised audio. The reason is that noise reduction will create distortion which ASR didn't see during training.

20 years from now, the future generation will laugh at how delusional some tech guys think that "text generation could be and end to humanity".

You can buy a USB microphone array and create one yourself. - https://pysdr.org/content/doa.html - https://www.minidsp.com/products/usb-audio-interface/uma-16-...

Based on credits in Gpt3 and 4 papers, I think the team that follows Sam and Greg are the main drivers of the tech. Ilya is an advisor more or less.

Go read the gpt3 and gpt3 tech report and see for yourself.

What motivates Sam to join Microsoft? Could it be the proximity of achieving AGI and his desire to remain competitive without starting from scratch?