Nicholas Colb

How to Use GPT for Generating Creative Content with Hugging Face Transformers Read this concise tutorial to find out how to use GPT to generate creative content with Hugging Face Transformers. No nonsense, just that facts.

6

Silicon Canals - European Technology News reports on Lovable, the Stockholm-based company making waves with its AI-driven GPT Engineer. The company has raised €6.8M in a pre-seed funding round led by Hummingbird Ventures and other notable investors including Shopify's Siavash Ghorbani and Creandum’s Stefan Lindeberg. Lovable, co-founded by Anton Osika and Fabian Hedin, aims to democratize software development for everyone – not just coders. Find out more about how this round will help push the boundaries of AI-driven software engineering. https://lnkd.in/eMwA3UpZ #AI #ArtificalIntelligence

2

10 AI Buzzwords You Need To Know in 2024: 1. AI Engineer: A Python engineer who uses requests library a lot. 2. Hallucinations: When ChatGPT acts like a stereotypical white male. Confident? Yes, but competent? Not always. 3. AutoML: I don't want to tune models manually anymore. 4. Guardrails: Time to prevent lawsuits from deploying poorly trained models… 5. Fine-tuning: Let’s spend our entire budget to tweak a model a bit. What could possibly go wrong? 6. Zero-Shot Learning: Passing an exam without ever seeing the questions. 7. Knowledge Graphs: We've connected all the dots. Now, if only our LLM could follow them… 8. AI Ethics: The topic we all pretend to care about but secretly find it to be the least exciting. 9. Vector Database: Because regular databases are too square for our high-dimensional problems. 10. RAG: A newborn golden baby that makes investors spend more money on LLMs. ... Add more in the comments 😝

151

32 kommenttia

NEW: “No blinks” mode in Calendar 🛠️ Use the Show recurrences toggle to hide link recurrences, aka link blinks 🔥 When hidden: 👉 Only the first appearance of new domains/links is shown 👉 Recurring disappearances are ignored 👉 Considered lost if domain/link does not reappear

29

2 kommenttia

LLM’s are bad for coders. And I’m not talking about junior coders only. Before we would copy paste code from Stackoverflow. It had rudimentary, but actually pretty good system for quality: votes. And you could not copy architecture or larger structures from there. Now we copy paste code from LLMs, which is complete and utter rubbish. 1. We copy paste bugs we don’t notice 2. We copy paste security flaws we don’t notice 3. We copy paste code which hinders performance of systems 4. We go back and forth copy pasting code that almost works, wasting a lot of time 5. We use it to “learn” new platforms without really actually learning them, compounding the wasted time effect 6. We get sluggish on our own thinking I like auto completion and this part works quite well these days. It’s also a good sparring partner in some cases. But copying entire code segments have almost always lead to time loss in my case. Am I doing something wrong?

15

18 kommenttia

Spent .9 hours debugging a missing <​meta name="viewport" content="width=device-width, initial-scale=1.0"> tag today—an elusive issue that even ChatGPT didn’t catch. As a software engineer focused on testing various LLMs/OpenAI models, these moments highlight the importance of human oversight in AI-driven development. It’s crucial to remember that AI is not infallible; it’s a tool that complements our skills. By identifying and addressing such bugs, we provide essential feedback for language model refinement, pushing the boundaries of what AI can achieve. — Let's embrace the potential of AI and evolve together, rather than fearing its capabilities. The future of tech lies in this symbiotic relationship, where human expertise and AI innovation drive unprecedented growth. #AI #SoftwareEngineering #TechInnovation #MachineLearning #OpenAI

26

OpenAI’s Whisper Speech-to-Text model can now be accessed 6X faster and 12X cheaper than on OpenAI’s API. This means all 25 hours of the Star Wars movies can be transcribed in ~7 minutes and ($0.5 per 1k minutes) and for ~$0.75 (that’s right, less than $1). Groq has just released public access to their Whisper API and fal has continued to improve performance of their Wizper offering which serves Whisper Large v3 using their custom inference engine. Faster and cheaper transcription opens up new use-cases for Speech-to-Text including for customer service and consumer use-cases which require a low cost-to-serve and responsive interactions. See below comment for link to our analysis:

75

2 kommenttia

From OpenAI Spring Update: ChatGPT is now available as a desktop app. Pretty cool, I can use that. New flagship model: GPT-4o is OpenAI’s new model and will be faster than previous models. It will also better integrate voice, text, and vision (o is for Omnimodel it seems). Available via the API as well. Improvement also with memory, already announced a while ago. 2x faster and 50% cheaper than GPT-4 turbo. Sounds like a good deal! Voice interface of ChatGPT can teach us to breathe, in case we forget. Live video analysis is pretty nice — gives a sudden feeling of conscious “presence”. If you show it “I ♥️ ChatGPT” it goes “awwww”.

3

Adding AI to a late software project only makes it later (the mythical AI-month) I'm bullish on LLMs for coding, but just because they type faster doesn't mean they 10x your output. The lessons from The Mythical Man-Month still ring true: - Communication isn't free: AI can streamline tasks, but it introduces new complexities. Aligning AI outputs with team goals requires coordination. - Division of labor isn’t simple: Tools like Devin overpromise. Breaking down complex requirements is hard. Creative problems need human judgment. - AI "onboarding" takes time: Integrating AI into workflows isn’t instant. Teams need time to learn AI tools and judge their outputs. - System integrity matters: AI-generated code can lead to inconsistencies. Maintaining a vision requires oversight. - Don’t set false expectations: Fast code generation leads to setting unrealistic deadlines. Real success lies in thoughtful integration. AI can improve software development, but execution speed usually isn’t the bottleneck.

40

11 kommenttia

people who say "with AI they can build anything" are DEAD WRONG..! here's my take: you do need to know atleast 50% of what real software engineering to use ai to build anything you want. if you don't know what the heck code editor or programming languages are then forget that you will ever be able to build anything complex. calculator apps—anyone can build but... you can't build the next slack if you don't know even the basics of software designing and development atleast as of now.... you need to have some skills in that niche to build something revolutionary with ai. in future too i think it'd get easy but.. most of it would be simplified by agents

2

🔄 𝐀𝐩𝐩𝐥𝐞 𝐖𝐚𝐧𝐭𝐬 𝐭𝐨 𝐌𝐚𝐤𝐞 𝐀𝐈 𝐔𝐩𝐝𝐚𝐭𝐞𝐬 𝐌𝐨𝐫𝐞 𝐒𝐞𝐚𝐦𝐥𝐞𝐬𝐬 Ever used a Large Language Model (LLM) and suddenly found your favorite prompt doesn't work anymore? You're not alone. This frustrating issue often arises after updates, which, while improving some aspects, can disrupt previously effective tasks. 𝐖𝐡𝐚𝐭'𝐬 𝐭𝐡𝐞 𝐏𝐫𝐨𝐛𝐥𝐞𝐦? Developers frequently fine-tune LLMs to reduce errors like hallucinations and glitches. However, each update can introduce new inconsistencies, causing LLMs to struggle with tasks they previously handled well. These "negative flips" occur even when training methods remain consistent across versions. 𝐀𝐩𝐩𝐥𝐞'𝐬 𝐒𝐨𝐥𝐮𝐭𝐢𝐨𝐧: Apple researchers are tackling this challenge head-on. They've developed metrics to identify differences between LLM versions, such as Meta’s Llama 1 and Llama 2. Using this data, they trained a specialized "compatibility model" to detect and flag discrepancies, reducing inconsistencies by about 40%. 𝐖𝐡𝐲 𝐈𝐭 𝐌𝐚𝐭𝐭𝐞𝐫𝐬: This research is a significant step toward smoother transitions between AI model updates. It highlights the importance of user experience in AI development and shows that tech companies are focusing on minimizing disruptions caused by updates. Stay tuned as Apple continues to refine this approach, making AI tools more reliable and user-friendly. #AI #MachineLearning #TechUpdates #AppleInnovation #LLM #AIResearch #UserExperience #TechSolutions #AIConsistency #FutureTech Source: https://lnkd.in/ggRCBt2x Image source: Apple

5

Dear Sam Altman, ​ ​ We owe you six big thank yous for holding off on launching Sora for public use until it’s safe. ​ ​ We’re on the same page that with AI’s great potential come major risks. But like that sixth finger in an AI-generated image, they often slip by unnoticed. ​ ​ That’s why we sent you a pair of – forget five – six-fingered gloves: to remind you to align AI with human values when approaching highly intelligent systems, and to encourage ecosystem collaboration on it.​ Best regards,​ Meeri and Veera from Saidot​ The AI Governance Company​ #WhatsYour6thFinger​ ​ #SamAltman #OpenAI #ResponsibleAI #AIsafety​ OpenAI

14

1 kommentti

What if you could fine-tune LLMs with chat commands like "Hey, I want to fine-tune LLaMA 3.1 405B". 🤔 Well, you don't have to imagine. During our recent visit to "The Fifth Elephant" event in Bangalore, programmers from different companies had several concerns about LLM Fine-tuning, the common ones included: • I know nothing about fine-tuning LLMs, but I know its complicated. • Fine-tuning LLMs is such a headache, I want something easier. • I tried some LLM fine-tuning platforms, but it doesn't make the job easy. That's where we come in. The MonsterAPI team has developed an AI agent that lets you fine-tune LLMs 10X faster with simple chat instructions. Say hi to "MonsterGPT". All you need to do is go to MonsterGPT, by clicking on this link here: https://lnkd.in/guCkJ8Ur and type "Hey, I want to fine-tune LLaMa 3.1 405B Here are some highlights of MonsterGPT: 1. No Coding Required: Perfect for non-technical users. 2. Comprehensive Support: Automatic hyperparameter tuning, memory efficiency, and scalability. 3. User-Friendly Interface: Seamless process with detailed explanations. 4. Deploy with Ease: Set up private API endpoints for your applications. Give it a try yourself. Interested in MonsterGPT? Read more about it here: https://lnkd.in/gUDc5Fj4

8

2 kommenttia

How does AI Language Models compare against each other? AI language models have revolutionized how we interact with technology, enabling sophisticated natural language processing, creative content generation, and complex problem-solving. In this performance evaluation by Lars Wiik, he tested the English language capabilities of OpenAI's GPT-4o, GPT-4 turbo, and GPT-4 against Google's Gemini 1.5 and Gemini 1.0. Using a custom dataset of 200 sentences across 50 topics to test classification performance, he evaluated these AI language models. The task was to match each sentence with the correct topic, and the results show that GPT-4o achieved the lowest error rate with only 2 mistakes. GPT-4, and Gemini 1.5 also performed strongly, each with just one more mistake than GPT-4o. Interestingly, GPT-4 Turbo did not perform as well as GPT-4, contrary to expectations. Gemini 1.0 lagged behind, which aligns with its lower price point. These findings highlight GPT-4o's superior performance in topic classification tasks. Choosing the right language model can significantly impact the success of your projects. Join us as we explore the advancements in AI and their implications for the future. Read more: https://lnkd.in/d_xts5vu #Ifelse #RevtraPro #Digijob #DigitalTransformation #InformationTechnology #AR #VR #UI #UX #Dev #ITStaffAugmentation #ITOutsourcing

1

Unpopular opinion: RAGs are not dead. They are probably your best shot. Google's Gemini Pro 1.5 has a 2M token window. You might be tempted to drop RAGs, and put everything in the context window. I think this is a mistake. And here is why: 1. "Needle in the haystack" ↳ Hidden info in long contexts gets lost. ↳ Recency bias might affect LLMs. 2. Price and performance ↳ Querying big models is expensive. ↳ Latency can be an issue. 3. Auditing is tougher ↳ Harder to track used information. ↳ Less transparency in relevance. RAGs offer more flexibility. LangChain's LangGraph builds a defined graph. That's how you can use: 1. Smaller LLMs efficiently. 2. Ensure high-quality output. 3. Have high visibility into performance. Ultimately you should find a good balance. Credits to Lance Martin for the summary. #RAG #LLM #ArtificialIntelligence #Data #ML

33

7 kommenttia

LLMs are destroying the lower end of developer jobs market. There's no sense in hiring juniors for dev jobs. What used to be a junior job, can be done now with Claude Sonnet by a non-tech person. This is great for the new generation of grads who are willing to learn the new toolset and get to the level of mids straight out of uni. Not great for people not willing to learn. Job market will be tough for them. Embrace #ai now.

19

3 kommenttia