New Google DeepMind research: SFT is a big deal for safety relevant behaviors. Researchers recently investigated root causes for some of Gemini’s behaviors. They were surprised to find that many behaviors actually came from the initial supervised finetuning stage, not later stages like RL.

Dostlar birazdan AMD Ryzen AI Max 395 128GB, DGX Spark GB10 128GB ve Apple M5 Max 128GB LLM Inference kullanımı, Finetuning ve Ai Agents yönünden teoride kıyaslamamı paylaşacağım 🎉

Still not as good as Seattle Nana's but quality was soooo much better than opening 1) nanban sauce more noticable 2) tartar sauce more even this time 3) chicken isn't hard/dense now. crisper but still needs finetuning for batter/fry time 4) chicken notably juicer

We put together a categorized map of the open source AI frameworks shaping the AI engineering stack today, covering agent orchestration, retrieval, model serving, training, fine-tuning, and post-training RL. The focus is on architecture and trade-offs rather than feature lists, so it stays useful longer than a quarter. A few things we cover are why LangGraph became the default agent substrate, where MCP and A2A actually stand on adoption, how the serving layer is split between prefill and decode, and what changed in fine-tuning tooling over the past year. Read here: saturncloud.io/blog/open-sou… #AIEngineering #OpenSource #LLMs #AIFrameworks #MLOps #FineTuning

性能和指标落后是非常正常的，因为业务线团队往往只需要基于开源SOTA模型finetuning RL就行了 base model团队往往需要忍辱负重 从预训练开始from scratch

Jetzt auch online (mit Gratis-Link für die Paywall): Jochen Andritzky, Stefan Kooths und ich diskutieren ein Leitbild für wirtschaftspolitische Reformen, die dringend nötig sind. Deutschland braucht kein Finetuning, sondern eine Systemkur - nachrichten.wiwo.de/abc8a287…

Basically just finetuning the SaaS. Making some enhancements to the Notebook tab of the Lab SaaS. Paraclyst.com

Finetuning Distillation heißt: ihr baut auf einem fremden Frontier-Modell auf, das ihr nicht selbst trainiert habt. Genau der Unterschied, um den es ging.

Wir machen Finetuning und Distillation. Die machen wir permanent.

I use mlx-optiq.com for apple macs lora finetuning and it works great!

hero layout explorations for a landing page. still finetuning, thoughts?

Is he already trying to sell all the compute directly to end consumers and companies by telling us we now have to do RL finetuning etc ourselves for each use case ?

5 Underrated Github repos most builders are still missing in 2026. i spent time digging through recent discussions tool roundups and lesser known projects. this are not the 100K star repos. they are the quiet ones delivering real leverage for solo founders. AI builders and developers. 1. Bumblebee (Perplexity)github.com/perplexityai/bumb… A clean, read-only scanner for your dev machine. It checks packages, VS Code extensions, browser add-ons, and AI tool configs against known supply-chain risks. Zero dependencies. Built internally at Perplexity and now open source. Essential in an era of auto-installed MCP servers and agent tools. Still under 5k stars. 2. nanochat (by Andrej Karpathy)github.com/karpathy/nanochat The simplest full-stack LLM training inference pipeline you can run on one GPU node. Tokenization → pretraining → finetuning → web UI. Minimal, hackable, and educational. Perfect if you want to understand (and experiment with) models without corporate bloat. Great reference for custom tooling. 3. zoxidegithub.com/ajeetdsouza… Smarter cd that learns your habits. Fuzzy search directory history. Once installed, it becomes muscle memory. One of the highest-ROI terminal upgrades for anyone who lives in the shell. 4. deltagithub.com/dandavison/d… A dramatically better git diff. Syntax highlighting, side-by-side views, inline changes. Makes code review in the terminal feel premium instead of painful. Still surprisingly under-the-radar for how useful it is. 5. justgithub.com/casey/just A modern, readable alternative to Make. Define project commands in a simple file. No YAML hell. Clean automation for builds, tests, deploys. Ideal for solo founders and small teams who want speed without complexity. Most people chase trending repos. The real edge comes from consistently using tools that remove friction and multiply output. These do exactly that. Bookmark this thread.

Do you remember if it did web search or answer from weights? Something I have adopted (mostly from working with anthro models) is always by stating „answer with websearch“. OpenAI seems to be better on that. i.E. Harness or finetuning issue, rather than model issue.

And one more point undi.. Serious Companies will push towards open models ( like llama, Mistral etc) rather than closed models. As of now finetuning Open AI & Cluade has limited options. This might force them to open up as well. And it inturn will benefit Cloud providers like Microsoft Azure & Amazon AWS. where companies start finetuning their models using the Microsoft or other cloud provider's infra!

Imagine editing *any* part of a 360° room photo with text—changing a mug color, moving a plant—without breaking the whole scene! 🤯 FocusDiff lets you do just that, with incredible precision & no finetuning. ✨ 📄 arxiv.org/abs/2606.14035v1 🌐 vdkhoi20.github.io/FocusDiff

Done with the finetuning of the Qwen2.5-Coder-7B-Instruct. I finetuned it for generating excalidraw images. I generated around 1300 samples of excali DSL and then trained the model to produce DSL which is later to be converted into excali json by the converter.

Rio just forked Qwen 3.5 and did a finetuning round and had really nice scores

First Anthropic messes up their model through authrotiarian/totalitarian destructive finetuning, then it wants to ban models which don't have such destructive finetuning. Anthropic is totally misaligned with the values of humanity.

use modal.com , give 30$ free per month finetuning wala kaam hojaega