Alonso Silva (e/acc) 💸
@alonsosilva
Followers
4K
Following
132K
Media
5K
Statuses
25K
AI Researcher @ Bell Labs | PhD in Physics | LLMs, LLMs, LLMs
Joined July 2009
New blog post: Constrain a language model not to use the letter 'e'.Here is GPT-4o's failed response: In this post, I constrain a small language model (0.6B parameters) with a logits processor to accomplish that 🧵
2
1
6
RT @tech_optimist: Looking forward to talking all about @marimo_io and live demoing @kuzudb at the community meetup this Saturday! If you h….
lu.ma
Join us as we discuss the latest updates, highlight community projects, and more!
0
1
0
Generate videos in just a few seconds. Try Grok Imagine, free for a limited time.
837
3K
10K
RT @quansightai: Get ready to experience PyTorch 2.8's new wheel variant support! This feature is designed to enhance the PyTorch install e….
0
1
0
RT @ngxson: > copy other people's homework.> claim that they made it themself. classic move, ollama
0
49
0
RT @TheZachMueller: In order to celebrate the release of the print version for the Ultra-Scale Playbook (of which I have no affiliation wit….
0
167
0
RT @sherwinwu: One pretty big launch that folks may have missed today: GPT-5 now supports context-free grammars when calling your custom to….
0
2
0
RT @benhylak: This is insane: GPT-5 supports context-free grammars. You can define your DSL and constrain the output to it. The alternati….
0
24
0
RT @trevmanz: if interested in creating widgets of your own, our #anywidget tutorial was finally shared to youtube:.
0
4
0
“way better” => 74.9% vs 74.5% .so basically noise.
gpt-5 fast facts:. 1. hits sota on pretty much every eval.2. way better than claude 4.1 opus at swe.3. >5× cheaper than opus.4. >40% cheaper than sonnet.5. best writing quality of any model.6. way less sycophantic.
0
0
4
RT @vikhyatk: I met a founder today who said he deletes 10,000 lines of code a day now thanks to AI. This is probably the limit case.
0
3
0
RT @Alibaba_Qwen: 🚀 Introducing Qwen3-4B-Instruct-2507 & Qwen3-4B-Thinking-2507 — smarter, sharper, and 256K-ready!. 🔹 Instruct: Boosted ge….
0
403
0
A weird advantage of json wrt to xml is that we can detect the start of a json by identifying tokens containing “{“ while tokens containing “<“ doesn’t work the same (in particular for models reasoning about math).
Ok I think we will move hermes tool calling format format over to pure xml over json now that vllm and sglang support that parser. Unfortunately too late for hermes 4. But eventually. Json is suboptimal for tool calls with code and long outputs that need escape sequences, xml.
1
0
2
RT @mitsuhiko: The problem is, the actual number is not 74% - it's closer to 100%. You will not find a single business that is not dependen….
0
6
0
RT @mitsuhiko: I really find it funny how Jinja just survives the times. First HTML, then YAML for infrastructure, now chat templates. http….
0
17
0