Data Machina #226
AI Coding Assistants. DeepSeek Coder. GPTutor. Fauxpilot. Tabby. Phind beats GPT-4. Videocrafter1. An AI Scientist. BloombergGPT. HelixNet.
On AI Coding Assistants. I’ve been meeting with a few companies that are exploring embedding AI coding assistants in their s/w dev pipelines. I’ve attended some fascinating conversations on the pros & cons of AI coding assistants, and also listened to some big political battles driving the AI agenda in these companies.
I read in the news that AI Job Openings Dry Up in UK Despite Sunak’s Push on Technology. Apparently, data from Reed Recruitment (one of the largest UK recruiters) shows postings linked to AI have dropped faster than for other roles. Perhaps UK companies are a bit more cautious about adopting AI? Meanwhile in the US, large VC firms are funding AI projects to bring autonomy to software engineering.
AI will replace/ won’t replace my coding skills. There are three camps here: 1) The Sr. managers who have no clue about AI coding assistants but think they can “remove some s/w engineers and reduce costs with AI” 2) Some old guard coding veterans who say “AI will never replace my coding skills I acquired in 20 years” and 3) Some enthusiastic engineers who are embracing AI for absolutely everything: “AI will empower my career…” How do you balance all the requirements for these 3 camps? It’s very tricky. In the meantime, here a few notes on AI coding assistants:
Building your own AI coding assistant. In this blogpost, the team at Deepsense.ai, share some of the challenges and limitations when developing an AI coding agent, and also share their insights and lessons learned from their experience. Blogpost: Creating your own code writing agent. How to get results fast and avoid the most common pitfalls.
Open-source alternatives to Copilot. There are very few open-source alternatives to Copilot. Here are four of the best open-source alts to Copilot:
Fauxpilot. An open-source locally hosted AI coding assistant. It uses the SalesForce CodeGen models inside of NVIDIA's Triton Inference Server with the FasterTransformer backend.
Tabby is a self-hosted AI coding assistant, offering an open-source and on-premises alternative to GitHub Copilot. Several key features include: 1)Self-contained, with no need for a DBMS or cloud service 2) Supports OpenAPI interface, easy to integrate with existing infrastructure (e.g Cloud IDE) 3) Supports consumer-grade GPUs.
GPTutor. A few weeks ago, researchers at CMU & Bucketprocol released a new open-source AI pair programming tool, as an alternative to GitHub Copilot. GPTutor empowers users to customize prompts for various programming languages and scenarios, with support for 120+ human languages and 50+ programming languages.
Cody (beta). This one is still in beta and has a freemium model. The development team at Sourcegraph, claim that Cody is “ the only AI coding assistant that knows your entire codebase.” Cody answers technical questions and writes code directly in your IDE, using your code graph for context and accuracy.
Beating GPT models at coding, program synthesis. A few notes on the very latest, new models outperforming GPT models at coding.
Phind Model beats GPT-4 at coding. This new model matches and exceeds GPT-4's coding abilities while running 5x faster. The model is built on top of open-source CodeLlama-34B fine-tunes, and has been fine-tuned on an additional 70B+ tokens of high quality code and reasoning problems.
DeepSeek Coder: State of the Art, open source. The latest SOTA performance among open code models. DeepSeker Coder is a series of code language models pre-trained on 2T tokens over more than 80 programming languages. Various model sizes (1.3B, 5.7B, 6.7B and 33B.) All with a window size of 16K, supporting project-level code completion and infilling. Open source and free for research and commercial use.
Concerns about AI Coding assistants. Beyond the common theme of “AI coding assistants generate productivity gains,” the fact is that many s/w engineering teams are reasonably concerned about the many potential issues around the embedding of AI coding assistants in their dev pipelines.
On the Concerns of Developers When Using GitHub Copilot This is an interesting new paper. A group of AI researchers from several unis, collected data from 476 GitHub issues, 706 GitHub discussions, and 184 Stack Overflow posts involving Copilot issues. The researchers identified the main issues, causes that trigger the issues, and solutions that resolve the issues when using Copilotjust.
Generate and Pray: Using SALLMS to Evaluate the Security of LLM Generated Code. Although LLMs can help developers to be more productive, prior empirical studies have shown that LLMs can generate insecure code. In this new, interesting paper researchers describe SALLM, a framework to benchmark LLMs' abilities to generate secure code systematically.
Have a nice week.
10 Link-o-Troned
the ML Pythonista
Deep & Other Learning Bits
AI/ DL ResearchDocs
data v-i-s-i-o-n-s
MLOps Untangled
AI startups -> radar
ML Datasets & Stuff
Postscript, etc
Tips? Suggestions? Feedback? email Carlos
Curated by @ds_ldn in the middle of the night.