![Elvis Saravia](/img/default-banner.jpg)
- 92
- 342 617
Elvis Saravia
United Kingdom
เข้าร่วมเมื่อ 5 ก.ย. 2013
Weekly tutorials, paper summaries, and technical walkthroughs on the most important AI and LLM developments.
Business inquiries: elvissaravia@dair.ai
Business inquiries: elvissaravia@dair.ai
Function calling with GPT-4o | OpenAI Playground
A demo on how to use function calling with the OpenAI Playground.
More on function calling in our guide: www.promptingguide.ai/applications/function_calling
Check out my upcoming live training to learn more about building with LLMs:
maven.com/dair-ai/prompt-engineering-llms
#ai #chatgpt #artificialintelligence
More on function calling in our guide: www.promptingguide.ai/applications/function_calling
Check out my upcoming live training to learn more about building with LLMs:
maven.com/dair-ai/prompt-engineering-llms
#ai #chatgpt #artificialintelligence
มุมมอง: 505
วีดีโอ
Scaling synthetic data, Agentless AI coder, RAG best practices | Top AI Papers of the Week
มุมมอง 277วันที่ผ่านมา
Summary of some of the most interesting AI and LLM papers of the week. 00:00 Tiny Giant - x.com/SFResearch/status/1807811770267971984 01:22 Million Tiny Experts - x.com/omarsar0/status/1810389538340290724 02:30 1B Personas - x.com/omarsar0/status/1807827401122238628 03:53 Reasoning in LLMs - x.com/omarsar0/status/1810329294884741594 05:20 Best Practices in RAG - x.com/omarsar0/status/1808177231...
Prompting and evaluating with Anthropic LLMs just got easier and faster!
มุมมอง 852วันที่ผ่านมา
Taking the new Anthropic Console prompting features for a test drive. Check out my upcoming live training to learn more about building with LLMs: maven.com/dair-ai/prompt-engineering-llms #ai #machinelearning #datascience #science #engineering
[LLM News] Moshi, RAG Best Practices, State of AI Report, Million Tiny Experts, GPT4All, RouteLLM
มุมมอง 696วันที่ผ่านมา
Another exciting episode of LLM News! Links mentioned in the video: 00:00 Moshi - x.com/kyutai_labs/status/1808883086173569222 01:54 Gen-3 Alpha - x.com/runwayml/status/1807822396415467686 02:19 RouteLLM - x.com/lmsysorg/status/1807812671238258931 04:10 Tiny Giant - x.com/SFResearch/status/1807811770267971984 05:23 Million Tiny Experts - x.com/omarsar0/status/1810389538340290724 06:31 1B Person...
Using LLMs to build a defense against adversarial attacks
มุมมอง 511วันที่ผ่านมา
Evaluates LLMs when used as a defense against adversarial attacks. Paper: arxiv.org/abs/2407.03234 Check out my upcoming live training to learn more about building with LLMs: maven.com/dair-ai/prompt-engineering-llms #ai #machinelearning #science #datascience
Jupyter Notebooks are now POWERED by AI (Codestral and GPT4-o)
มุมมอง 970วันที่ผ่านมา
Overview of Pretzel, a new AI-powered Jupyter. Repo: github.com/pretzelai/pretzelai The dataset used: github.com/dair-ai/ML-Papers-of-the-Week/blob/main/research/ml-potw-10232023.csv #ai #machinelearning #scienceandtechnology #coding #datascience
[LLM News] ESM3, CriticGPT, Gemma 2, LLM Compiler, LongRAG, GraphReader
มุมมอง 1.3Kวันที่ผ่านมา
Another exciting episode of LLM News! Links mentioned in the video: 00:00 ESM3 - th-cam.com/video/l2EvtcLb19o/w-d-xo.html&ab_channel=ElvisSaravia 01:13 Gemini Announcements - developers.googleblog.com/en/new-features-for-the-gemini-api-and-google-ai-studio/ 03:32 Gemma 2 - th-cam.com/video/vJyCr2yPTQM/w-d-xo.html&ab_channel=ElvisSaravia 05:26 ChatGPT on Desktop - openai.com/chatgpt/mac/ 06:19 C...
Google releases Gemma 2 and it's IMPRESSIVE!
มุมมอง 2.9K14 วันที่ผ่านมา
Reviews the latest updates on Gemini 1.5 models and experiments with the new Gemma 2 model. 00:00 New Gemini announcements 03:01 Long-context & context caching 04:01 Code Execution 06:01 Gemma 2 Testing Official Gemini Announcement: developers.googleblog.com/en/new-features-for-the-gemini-api-and-google-ai-studio/ Context caching tutorial: th-cam.com/video/987Pd89EDPs/w-d-xo.htmlsi=TRaG-3ToghQ1...
AI discovers NEW green fluorescent protein!
มุมมอง 47314 วันที่ผ่านมา
A summary of the new ESM3 paper by Evolutionary Scale. More here: www.evolutionaryscale.ai/blog/esm3-release Paper: evolutionaryscale-public.s3.us-east-2.amazonaws.com/research/esm3.pdf #science #ai #biology
Few-Shot Prompting Explained
มุมมอง 2.5K14 วันที่ผ่านมา
In this video, I explain the idea behind few-shot prompting, what enables it, and how it can be used with LLMs. More in our guide: www.promptingguide.ai/techniques/fewshot Learn more about few-shot prompting and LLMs in one of our upcoming live courses: maven.com/dair-ai/prompt-engineering-llms #ai #llms #promptengineering #machinelearning #programming
[LLM News] Claude 3.5 Sonnet, Open-Sora, Context Caching, PlanRAG, Safe SuperIntelligence Inc
มุมมอง 1.1K21 วันที่ผ่านมา
Another exciting episode of LLM News! Links mentioned in the video: 00:00 Claude 3.5 Sonnet - www.anthropic.com/news/claude-3-5-sonnet 04:04 Gen-3 Alpha - runwayml.com/blog/introducing-gen-3-alpha/ 05:05 Safe SuperIntelligence Inc - x.com/ssi/status/1803472825476587910 06:35 Meta AI Research - about. news/2024/06/releasing-new-ai-research-models-to-accelerate-innovation-at-scale/ 07:49 De...
New Claude 3.5 Sonnet is here and it's POWERFUL!
มุมมอง 4.1K21 วันที่ผ่านมา
New Claude 3.5 Sonnet is here and it's POWERFUL!
MoME Reduces LLM Hallucinations by 10X!
มุมมอง 9K28 วันที่ผ่านมา
MoME Reduces LLM Hallucinations by 10X!
[LLM News] Apple Intelligence, Dream Machine, Self Teaching, AI Personality, Function Calling Course
มุมมอง 440หลายเดือนก่อน
[LLM News] Apple Intelligence, Dream Machine, Self Teaching, AI Personality, Function Calling Course
Function Calling with OpenAI APIs | A Crash Course
มุมมอง 3.6Kหลายเดือนก่อน
Function Calling with OpenAI APIs | A Crash Course
[LLMS News] AGI Predictions, Mamba-2, NLLB, GPT-4 Features, Structured LLM Generation, KLING
มุมมอง 921หลายเดือนก่อน
[LLMS News] AGI Predictions, Mamba-2, NLLB, GPT-4 Features, Structured LLM Generation, KLING
New prompting method uses thought templates | Buffer of Thoughts
มุมมอง 1.2Kหลายเดือนก่อน
New prompting method uses thought templates | Buffer of Thoughts
Extracting features from Claude 3 Sonnet
มุมมอง 342หลายเดือนก่อน
Extracting features from Claude 3 Sonnet
[LLM News] xAI Series B, Codestral, LLM Guide, AutoGen Course, Symbolic Chain-of-Thought
มุมมอง 999หลายเดือนก่อน
[LLM News] xAI Series B, Codestral, LLM Guide, AutoGen Course, Symbolic Chain-of-Thought
Exploring Capabilities of Long-Context LLMs
มุมมอง 598หลายเดือนก่อน
Exploring Capabilities of Long-Context LLMs
[LLM News] GPT4-o, Project Astra, Veo, Copilot+ PCs, Gemini 1.5 Flash, Chameleon
มุมมอง 795หลายเดือนก่อน
[LLM News] GPT4-o, Project Astra, Veo, Copilot PCs, Gemini 1.5 Flash, Chameleon
[LLM NEWS] KANs, Gemma 10M Context, OpenAI Updates?, Automatic Prompt Engineering, Tokenizer Arena
มุมมอง 7002 หลายเดือนก่อน
[LLM NEWS] KANs, Gemma 10M Context, OpenAI Updates?, Automatic Prompt Engineering, Tokenizer Arena
[LLM NEWS] AlphaFold 3, xLSTM, OpenAI's Model Spec, DeepSeek-V2, OpenDevin CodeAct 1.0
มุมมอง 1.5K2 หลายเดือนก่อน
[LLM NEWS] AlphaFold 3, xLSTM, OpenAI's Model Spec, DeepSeek-V2, OpenDevin CodeAct 1.0
SWE-Agent | An LLM-based Software Engineering Agent
มุมมอง 1.1K2 หลายเดือนก่อน
SWE-Agent | An LLM-based Software Engineering Agent
Better and Faster LLMs via Multi-token Prediction
มุมมอง 2.2K2 หลายเดือนก่อน
Better and Faster LLMs via Multi-token Prediction
Training an LLM to effectively use information retrieval
มุมมอง 1K2 หลายเดือนก่อน
Training an LLM to effectively use information retrieval
amazing video and great resources! Thank you Elvis! <3
This is very good, thank you for your effort!
Thanks for the support
this is a great format. your input is really beneficial. only thing is if we can break it into two videos, each no more than 10 minutes, instead of one long 20 minutes. it's a preference, not vital.
Appreciate the feedback. Yes, I agree that it makes sense to potentially split it or significantly shorten it. Maybe doing the episodes twice a week might be better. Thanks.
It is a good paper, but unfortunately use Llama guard, retrieval, rerank, etc. Will be LLMs slower in latency with a lot of inferences, those are the only disadvantages that I see.
Those are good points. I think they mention a few of these things in the discussion. Smaller language models may be more ideal here.
Thanks for sharing
amazing video
how do I get more tokens
Thx for AI news update, really appreciate it
Great!
first comment 🤣🤣🤣🤣
:)
Nice
Thanks
What up my dude! Thanks
Thanks, my dude!!
As always, a good value add summary, thx
Where is the context cashing?
You can find the link to a separate tutorial I did for the context caching in the description.
Great Information! Keep it up
Elvis news video series?! Let's gooo!
is strongly censored, he doesn't even give advice on how to solve a headache!
It does seem that there are strong safety guardrails on this one.
Cool
i like your explainer on Twitter, but i like youtube explainer more!!
Is gemma 2 good for coding? Like is it better than DeepSeek Coder V2?
It's good for coding but I don't think it's better than a custom model like DeepSeek Coder V2. I think DSCV2 is a lot better.
I didn't understand the candle task too. Maybe, I'm also an AI after all...
It's meant to be a puzzle but logical reasoning is required to get this one right.
Its funny because because for developers and longer windows. 2:40 Does Interpreter execute code or just count the delimiters or -> like: if the code included a function requiring advances read/write permission when executing interpreter would write to file -> ?: Is what I thought? 3:15 I think cache could be leveraged in place of compute while multiprocessing but per thread/semaphore.
Cool
Thank you for sharing this
Its fantastic at writing fiction but it hallucinates when asking it detailed IT questions. I used it for a few days and noticed the difference with GPT 4o in capablility. I canceled my subscription to gpt 4 and subscribed to Claude but ended up changing back.
Thanks for sharing. Yeah, I have seen the challenged with IT or technical question. This is a test I always run before committing to any LLM because most of my uses cases involve technical content. I haven't cancelled any of my subscriptions. Still experimenting a lot with the tools.
is already available in VertexAI Tool?
Not checked yet.
Thanks, awesome video. What prompt did you use for the MLP? Wasn't able to recreate it but I destpreatly need as I wanna do an introductory lecture series on deep learning
I am not sure it’s possible to share the full prompt from Claude. Let me look into it or I might do a video explaining the process.
All just to say "Those who control history controls the future" and "You will own nothing and be happy".
Ooh, is it out on llama?
Yes ollama has it. Just make sure to update to the latest version as it has important bug fixes. 1.46 as of now
Thanks for responding.
Can you have it write you some new music?
Doubt it has that capability but it could be interesting
This is awesome. We need more updates and explanations like this!
What application you used to read pdf paper? Is that Adobe reader?
It’s just Mac preview
God is good, these tools were all included and better in our 3.5 watt brains 🧠
Claude need to implement a differ. Regenerating full code of artifacts evey time, even for minor changes is so inefficient.
I don't understand how it was 5 times more expensive and 2x slower three to four months ago when it was only 3 Opus, now Sonnet is 3.5 and is even more intelligent than Opus, faster and cheaper in just 3 to 4 months.
They figured something. Same happened for the newer openai models. They got cheaper and better.
@@elvissaravia that's fascinating!
3.5 sonnet has less context than 3 opus
7:45 8-bit Luigi looks more like Walter White 😄
😂 true! Maybe giving it an image as a reference could have helped
Let me know if you want me to do a separate video with more examples. I have a lot more examples that I am sharing here: x.com/omarsar0/status/1803796159334322566
it's awesome
Looks very intelligent
A question on this, In the tutorial above we are using that txt file to answer our queries, it's more like grounding the results of our query to certain datastore (txt file in this case). So what is the difference in terms of functionality for using this context cached feature and grounding response on some datastore (with this txt file stored in it). Appreciate response.
We don't know exactly how the cached context is used at inference but I assume it's not much different from the latter you mentioned.
Thank you,,,
Great stuff Elvis, just subscribed annually to your newsletter. :)
Thank you. I appreciate the support.
Can you please provide the code
Let me work on that. Check back by end of the day and I will share the link here.
Thank you so much for taking out time to explain this
Thanks!
Nice! That does add a lot more comfort in correct answers. The "mixture of agents" model architecture is coming in with some good stuff too (not as good as this though - this is big). We're not far from some really smart agents...
Hmm. Lots of PR stunts on their blog. So still... skeptical. I really don't get the main trickery, and 200 API calls per month is not enough to get a proper test-through. "Internal memorization. Tuning the weights, not RAG. You can layer them." /via X.
Loved the tone on the content man, you've got a new subscriber! Great job!
But is it right to call this as innovation ? Just training million of experts with task specific facts can't be said to be research ?
It’s special because it swaps in those experts within a larger architecture. Related research on polysemanticity also suggests that sparsity will enhance explainability and steer ability