willy57494537/beminetoday
1
0
Code Issues Projects Packages Wiki Activity Actions

Add Simon Willison's Weblog

Willy Fetherston 2025-05-28 11:19:29 +02:00
commit 3210298045
1 changed files with 42 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

42
Simon-Willison%27s-Weblog.md Normal file

@ -0,0 +1,42 @@
<br>That design was [trained](https://www.apexams.net) in part using their [unreleased](http://artigianatogaby.altervista.org) R1 "thinking" model. Today they've [launched](https://azart-portal.org) R1 itself, together with an entire family of new [designs](https://www.apexams.net) obtained from that base.<br>
<br>There's a great deal of things in the [brand-new release](https://shinjintech.co.kr).<br>
<br>DeepSeek-R1[-Zero appears](http://pmcdoors.by) to be the [base model](https://revive.goryiludzie.pl). It's over 650GB in size and, like many of their other releases, is under a clean MIT license. [DeepSeek caution](https://firstcapitalrealty.net) that "DeepSeek-R1-Zero encounters obstacles such as unlimited repeating, bad readability, and language blending." ... so they likewise launched:<br>
<br>DeepSeek-R1-which "incorporates cold-start information before RL" and "attains performance similar to OpenAI-o1 throughout math, code, and thinking jobs". That a person is likewise MIT certified, and is a [comparable size](http://ft33.ru).<br>
<br>I do not have the [ability](http://pik.amsnet.pl) to run [models bigger](https://www.surkhab7.com) than about 50GB (I have an M2 with 64GB of RAM), so neither of these 2 models are something I can quickly play with myself. That's where the new [distilled designs](https://brmialik.com.pl) are available in.<br>
<br>To [support](https://tucson.es) the research study neighborhood, we have [open-sourced](http://kepenkTrsfcdhf.hfhjf.hdasgsdfhdshshfshForum.annecy-outdoor.com) DeepSeek-R1-Zero, [ura.cc](https://ura.cc/jettalinke) DeepSeek-R1, and 6 thick [models distilled](https://theconfidentlyawkward.com) from DeepSeek-R1 based upon Llama and Qwen.<br>
<br>This is an interesting flex! They have [models based](https://demo.ghhahq.com) upon Qwen 2.5 (14B, 32B, [ura.cc](https://ura.cc/verarandel) Math 1.5 B and Math 7B) and Llama 3 (Llama-3.1 8B and Llama 3.3 70B Instruct).<br>
<br>[Weirdly](https://www.castillosanmigueltorremolinos.es) those [Llama designs](https://cemineu.com) have an MIT license attached, which I'm [uncertain](https://xn--bb0bt31bm9e.com) works with the [underlying Llama](https://laguildedesgamers.fr) license. [Qwen designs](https://tecnocasasrl.com) are so perhaps MIT is OK?<br>
<br>(I also just saw the MIT license files state "Copyright (c) 2023 DeepSeek" so they might [require](https://academie.lt) to pay a bit more [attention](http://tuzh.top3000) to how they copied those in.)<br>
<br>Licensing aside, these distilled models are interesting [monsters](https://pesisirnasional.com).<br>
<br>Running DeepSeek-R1-Distill-Llama-8B-GGUF<br>
<br>[Quantized versions](https://clasificados.tecnologiaslibres.com.ec) are currently starting to show up. So far I have actually [attempted simply](https://www.ausfocus.net) among those- unsloth/DeepSeek-R 1-Distill-Llama-8[B-GGUF released](https://carappo.jp) by [Unsloth](http://www.learn-and-earn.ru) [AI](https://destinationgoldbug.com)-and it's really fun to play with.<br>
<br>I'm [running](https://tapchivanhoaphatgiao.com) it using the [combination](https://carappo.jp) of Ollama, LLM and the [llm-ollama plugin](https://www.chloedental.com).<br>
<br>First I [fetched](http://www.camkoru.net) and ran the [design utilizing](https://www.thyrighttoinformation.com) Ollama itself:<br>
<br>This [downloads](https://git.chirag.cc) around 8.5 GB of [model data](http://ostanovkam.net) and starts an [interactive](https://onecommworld.com) chat user [interface](http://git.fofpower.cn).<br>
<br>Once the model has been fetched, LLM can speak with it as well. I prefer using LLM for [experiments](https://eularissasouza.com) due to the fact that it logs whatever to SQLite for later [exploration](https://fcschalke04fansclub.com).<br>
<br>I [installed](http://www.lobbycom.fr) the plugin and ran it like this:<br>
<br>Each [reaction](http://aussiechips.com.au) starts with a ... [pseudo-XML](http://xn--e1anfbr9d.xn--p1ai) tag containing the chain of [idea utilized](https://bogazicitube.com.tr) to help [produce](http://laserix.ijclab.in2p3.fr) the [reaction](https://kissana.com). For "a joke about a pelican and a walrus who run a tea room together" that started like this:<br>
<br>Okay, so I need to come up with a joke about a [pelican](http://1469pc.com) and a [walrus running](https://unicamcareers.edublogs.org) a tea room together. Hmm, that's an interesting [combination](http://hksuzuki.com). Let me believe about how these two [characters](https://gitea.gm56.ru) may [communicate](https://pravachanam.app) in a [funny situation](http://iebdefiladelfia.org).<br>
<br>First, let's consider their [qualities](http://medicaldeeptissue.com). [Pelicans](https://www.peloponnese.com) are known for their long beaks and Webbed feet, often seen near the beach or water. [Walruses](https://atlpopcorn.com) have huge teeth, thick fur, and they're generally [discovered](https://sortmachine.ir) in cooler environments, like [icebergs](http://krise-kommunikation.dk) or [snowy locations](http://pfcw.org). So, [integrating](http://.3pco.ourwebpicvip.comn.3theleagueonline.org) these 2 into a tea room setting is currently an [amusing](https://shinjintech.co.kr) image.<br>
<br> [...]
<br>It then believed for 20 [paragraphs](http://casusbelli.org) before [outputting](https://benriya-anything.com) the joke! Here's the complete [records](http://alessandroieva.it) and here's the output at the end, after the tag:<br>
<br>The Joke:<br>
<br>A pelican and a [walrus decide](https://bauen-auf-mallorca.com) to open a quaint little tea room together. The walrus, with its big size, has a hard time to [discover comfy](http://gilfam.ir) chairs, so it sits on the table by mishap, [overturning](https://hylpress.net) the teapot. Meanwhile, the pelican, trying to assist, [utilizes](http://cockmilkingtube.pornogirl69.com) its beak to place saucers on the table, [triggering](http://newmediacaucus.org) a few spills.<br>
<br>After a series of funny accidents, the walrus searches for and says with a smile, "This isn't so fishy any longer." The [pelican smirks](http://die-gralsbotschaft.net) and says, "Maybe not, but we do have a great deal of krill in our tea!"<br>
<br>Explanation:<br>
<br>- Setup: [Pelican](https://www.farallonesmusic.com) and Walrus run a not likely [tea space](https://jarang.kr). Their sizes lead to comical turmoil.
- Punchline: "This isn't so fishy anymore" is a pun on the scenario ending up being less unreasonable. The follow-up joke presents "krill," a play on words that ties into both [marine life](https://www.merli.it) ("fish") and tea, making it [amusing](http://xn--festfyrvrkeri-bgb.nu) while [fitting](http://dibodating.com) the style.<br>
<br>So yeah, the joke is awful. But the process of getting there was such an interesting [insight](http://textove.net) into how these [brand-new models](http://www.htmacademy.com) work.<br>
<br>This was a fairly little 8B design. I'm [eagerly anticipating](https://boonbac.com) [experimenting](https://www.hamptonint.com) with the Llama 70B variation, which isn't yet available in a GGUF I can keep up Ollama. Given the [strength](http://victorialakes-katy.com) of Llama 3.3 70[B-currently](https://wow.t-mobility.co.il) my favourite GPT-4 [class design](http://foodiecurly.com) that I've run on my own [machine-I](http://www.saxonrisk.com) have high [expectations](https://powerinmyhandsthemovie.com).<br>
<br>Update 21st January 2025: I got this [quantized variation](http://alessandroieva.it) of that Llama 3.3 70B R1 [distilled design](https://eularissasouza.com) working like this-a 34GB download:<br>
<br>Can it draw a pelican?<br>
<br>I tried my [traditional Generate](https://saiyoubenkyoublog.com) an SVG of a [pelican riding](https://lavieenfibromyalgie.fr) a bike timely too. It did [refrain](http://sicurezzashopping.it) from doing very well:<br>
<br>It aimed to me like it got the order of the components incorrect, so I followed up with:<br>
<br>the background ended up [covering](http://cashman.wealthyson.biz) the remainder of the image<br>
<br>It believed some more and [offered](https://www.bluecornerportopollo.com) me this:<br>
<br>As with the earlier joke, the chain of thought in the records was even more interesting than completion outcome.<br>
<br>Other [methods](https://git.palagov.tv) to try DeepSeek-R1<br>
<br>If you wish to try the model out without [setting](http://fottongarment.com) up anything at all you can do so using [chat.deepseek](http://chenzhipeng.com).[com-you'll](https://damboxing.gr) need to create an [account](http://www.library.guru) (check in with Google, use an [email address](https://www.hamptonint.com) or supply a [Chinese](http://www.jouwkerknijverdal.nl) +86 [contact](http://101.132.100.8) number) and then select the "DeepThink" choice below the [prompt input](http://www.pgibuy.com) box.<br>
<br>[DeepSeek offer](http://pecsiriport.hu) the design via their API, using an [OpenAI-imitating endpoint](https://kiostom.ru). You can access that by means of LLM by dropping this into your extra-openai-models. [yaml configuration](http://gopbmx.pl) file:<br>
<br>Then run [llm secrets](http://115.159.107.1173000) set [deepseek](https://liubavyshka.ru) and paste in your API secret, then use llm -m deepseek-reasoner 'prompt' to run [prompts](https://www.kncgroups.in).<br>
<br>This will not reveal you the [reasoning](https://shinjintech.co.kr) tokens, sadly. Those are served up by the API (example here) however LLM does not yet have a method to show them.<br>