PICK YOUR SUPPORT STYLE
MONTHLY SUPPORT
Reader
$5/mo
Contributor
$15/mo
Architect
$50/mo
Recurring subscriptions auto-bill monthly via Stripe Checkout. Cancel anytime from the receipt email.
What was talked about: Meta’s Muse series replaces the Llama name, with Muse Glimmer 2.1 positioned as a 30-billion-parameter dense model that can run on consumer hardware such as an RTX 3090. Its performance appears close to comparable Qwen and Gemma models, while Meta has reaffirmed that open source remains an important part of its strategy.
Takeaway: Muse Glimmer expands the choice of capable local models, but it does not yet offer a clear quality advantage over established alternatives.
Links shown: Muse announcement, Muse benchmarks and DFlash
What was talked about: DeepFlash speculative decoding can make local, batch-size-one inference roughly three times faster, with high-end hardware potentially reaching more than 200 tokens per second. Qwen 3.8 is expected to replace the current 27-billion-parameter model soon, while both Meta and Alibaba have announced plans to release much larger frontier-scale models openly.
Takeaway: Inference improvements make dense local models more practical, but upcoming Qwen releases may quickly surpass Meta’s current model.
Links shown: Muse announcement, Muse benchmarks and DFlash
What was talked about: Muse Glimmer is a general-purpose agent model with mixed coding results. Some reported coding benchmarks are considered weak indicators, while Terminal-Bench places it below Qwen 3.6 and SciCode shows stronger scientific programming performance. Early real-world reports describe it as competent but not distinctive.
Takeaway: Use stronger benchmarks and real workloads before changing models; Muse Glimmer is a credible option, not a necessary upgrade.
Links shown: Muse announcement, Muse benchmark table
What was talked about: A family-history application converts natural-language transcripts into stories suitable for printed keepsake books. It extracts people and relationships into a knowledge graph, identifies both explicitly told stories and narratives implied across generations, and then creates drafts for human review.
Takeaway: Structured entities and relationships can reveal meaningful narratives that a simple transcript summarization pass would miss.
Links shown: Narrative prose pipeline
What was talked about: The original story pipeline used Gemini to extract and write each story in one sequence. A new editing stage uses Opus 5 to revise the prose before human review. Although Opus is expensive per token, the short length of each story keeps the total editing cost small.
Takeaway: For short, high-value writing, a separate premium editor can improve quality at a manageable cost.
Links shown: Narrative prose pipeline
What was talked about: Blind head-to-head evaluation with an LLM judge produced 28 wins, three ties, and five losses for the edited pipeline. The rubric covered source grounding, unnecessary padding, missing or flattened details, voice, coherence, and overall readability. GPT Luna also produced editing results close to Opus 5 in the comparison.
Takeaway: A defined rubric and blind comparisons make prose improvements measurable and can identify lower-cost models with similar results.
Links shown: Narrative prose pipeline
What was talked about: Writer and editor combinations can have self-preference bias, so using separate models may expose weaknesses that self-editing misses. The production cascade uses Gemini Flash Lite for fast extraction, Gemini Flash for story detection, Gemini 3.1 Pro for writing, and Opus 5 for editing. Flash Lite also provides low latency and reliable structured JSON output.
Takeaway: Assign each model a narrow role based on speed, cost, and quality instead of using one expensive model for the complete workflow.
Links shown: Narrative prose pipeline, Gemini 3.5 Flash-Lite analysis
What was talked about: A future planning stage could organize the source facts and narrative shape before writing and editing begin. The broader product aims to capture family histories through voice or text and turn them into durable printed books, including adult keepsakes and illustrated children’s stories based on real family events.
Takeaway: A plan-before-write stage may improve narrative structure, while printed output gives generated family stories a durable purpose beyond the application.
Links shown: Narrative prose pipeline
What was talked about: Pi is a minimal coding-agent harness with only a few core tools and low token use. OMIPi extends that base with a larger tool set, abstract syntax tree search, language-server features, static analysis, definition lookup, and debugging. Its context graph can reduce search latency and token consumption while supporting self-hosted or open models.
Takeaway: Harness design can affect coding performance as much as model choice; efficient context retrieval often produces faster and cheaper results.
Links shown: Narrative prose pipeline, Pi repository, Pi website, oh-my-pi repository, OMP website
What was talked about: Prime Intellect’s harness uses a recursive language model approach in which a primary model creates sub-agents through an IPython environment and passes context through variables. The system can adapt its tooling to a benchmark, but its ARC-AGI results are controversial because the harness performs substantial reasoning outside the base model. Recursive calls also increase cost, infrastructure load, and the risk of uncontrolled delegation.
Takeaway: Recursive agents are promising for dividing context, but they need strict depth limits, clear roles, and bounded costs to remain reliable.
Links shown: oh-my-pi repository, Prime Agent announcement, Prime Agent repository, Lambda-RLM repository
What was talked about: OpenCode offers inexpensive access to open models and extensive workflow configuration, including separate models for planning, implementation, and review. Its default prompts and context management can be weaker than competing harnesses, causing high token consumption that becomes much worse with sub-agents. Custom prompts, pruning, and project instruction files can materially improve the experience.
Takeaway: OpenCode is most useful for teams willing to configure it; its flexibility comes with added tuning and token-management work.
Links shown: Prime Agent announcement, OpenCode repository, OpenCode
What was talked about: Powerful local coding agents introduce concerns about dependency supply chains, remote tools, API credentials, and privileged repository access. Open source permits direct auditing, but tool calls and external MCP servers remain possible attack surfaces. A security scan of an AI code-review tool also found prompt-injection opportunities through malicious pull-request comments.
Takeaway: Audit the harness, its dependencies, and every external tool boundary, and treat repository content as untrusted input to the agent.
Links shown: OpenCode repository, OpenCode
What was talked about: OpenAI’s Daybreak Blue and Red models target defensive security analysis and offensive red teaming, alongside a Codex security auditing tool. Separate research shows that hidden reasoning traces can be extracted by generating an answer with a strong model and then asking a weaker model in the same conversation to reveal the prior reasoning. The leaked traces expose benchmark memorization, unusual internal reasoning patterns, and potentially valuable data for model distillation.
Takeaway: Security models can improve code auditing, but prompt injection and cross-model reasoning leakage remain unresolved platform-level risks.
Links shown: GPT-5.6-Cyber and Daybreak post, Reasoning-trace extraction post, Stolen Thoughts paper
What was talked about: Anthropic plans to add statistical watermarks to Claude-generated text by subtly changing token selection rather than attaching removable metadata. The technique can survive limited edits but can be weakened by extensive rewriting or passage through another model. The broader provenance problem includes unresolved lawsuits over copyrighted training data. For nontechnical teams building an AI-agent MVP, n8n provides visual agent workflows and scalable execution, while Lovable is better suited to building a website and its interface.
Takeaway: Text watermarks provide evidence of origin but are not tamper-proof; choose workflow tools such as n8n for agent logic and website builders such as Lovable for the user interface.
Links shown: Reasoning-trace extraction post, Claude watermark report, n8n
0:02 All right, welcome back everyone to AI Tools Club.
0:06 Yeah, let’s get right into things. Thanks for coming back, as always.
0:11 We will start this week with a new model from Meta, actually open source model.
0:17 Not technically a new Llama model. They’ve changed their model series names to Muse instead.
0:22 But many people are calling this Llama 5.
0:25 And it is called, yeah, Muse Spark 2.1.
0:28 So this is based off their Glimmer.
0:34 MuseSpark is their big model. MuseGlimmer is the smaller open source one.
0:38 This is a 30 billion parameter dense model, actually.
0:41 It’s not a mixture of experts models like Quen35B.
0:45 It is more similar to Quen 3.627B. It’s meant to be run at home on devices like 3090.
0:54 Performance-wise, it’s probably about on par, maybe a little bit worse than the Gemma and Quenn models.
1:05 Yeah, but it’s nice to see that Meta is coming back to the open source space a bit.
1:12 Mark Zuckerberg actually wrote a big sort of like letter or I don’t know what else he would describe it, but or memo, basically describing that he sees open source as an important pillar for meta still.
1:26 And they actually plan on open sourcing their MuseSpark 2.1 model, like you said, the bigger model right now, that I believe, at least on artificial analysis, is around like Opus 4.8 level or
1:40 like Proc 4.5, somewhere around there.
1:45 Although in the real world, it doesn’t seem to be notable in any main way, which is why, yeah, we haven’t talked about it too much.
1:53 But yeah, no, this open source model, great to see.
1:56 This is sort of, it seems to be the size that a lot of these companies are aiming for.
2:01 It’s like I said, Quen, they’ve got models this size.
2:05 And then also Gemma, they also have a dense model around the same size as well.
2:10 They also have DeepFlash speculative decoding.
2:13 This is from DeepSeek, some DeepSeek research, but this is speculative decoding.
2:17 So you basically just turn this on and your model is three times faster at batch size one.
2:23 So the way you’d use it locally. So yeah, if you got a 50-90, you can be running this model at 233 tokens per second, which is pretty quick.
2:30 That’s about five times faster than GPT 5.6 Sol will be.
2:36 It’s yeah, it’s I haven’t been able to play with it too much myself.
2:41 If it’s around the Quen 3.6 level, that means it’s a fairly usable model, I would say.
2:47 I think people compare it usually somewhere between like Sonnet 4 and 4.5 levels of capability, which is you could definitely do stuff with it at home.
2:58 But I also don’t think this model is going to be relevant for more than about two more days because Quen has announced that they are going to be releasing Quenn 3.8 this week.
3:11 And so they said they’re doing the refresh for their 27 billion parameter dense model.
3:16 And so right now, like I said, this is already about equivalent to that, maybe a little bit worse.
3:20 So I expect Quen 3.8, 27b to sort of blow it out in the water.
3:25 So yeah, if you want to use this model, use it for two days or so, and then you will be disappointed and be using a worse model, most likely, versus the new Quenn model that’s coming out.
3:36 But yeah, cool to see this. Cool to see that Meta is staying in the open source game.
3:41 There was some worries there for a while that they wouldn’t be.
3:44 But they’re making a return. So hopefully we see more in the future because having more options is always a good thing for this sort of stuff.
3:54 Yeah, any questions about that? Do we know about the vibe of the model?
4:04 What is it good at? Is it a coding model?
4:08 It’s a general, like just general agent.
4:10 I mean, it benchmarks well on some of their coding benchmarks, I think.
4:16 But they’re not using any of the really good ones.
4:18 Like, SuiBench Pro and SuiBench Verified are two benchmarks that are known to be bad.
4:22 So we do ignore those. Terminal Bench is fairly decent, and it does worse than Quen 3.6.
4:28 Sci code, which is scientific coding, so like matplotlib kind of stuff, it does better there.
4:33 From the vibe text, I haven’t seen too many people using it.
4:36 It’s only been out for about 24 hours now, I believe.
4:40 I’ve seen some people using it and saying it’s decent, but if you’re already running Quen 3.6 or any of these other models, there’s probably no good reason or any real need to go and use it.
4:51 Like I said, it’s a competitor. It’s cool to see meta in the space, but I don’t think it’s game-breaking or sort of unique in any way that the other models aren’t already.
5:03 Well, it’s unique. It’s not an utter slop.
5:06 It’s yeah, I haven’t heard any bad things about it per se.
5:11 Where it’s, you know, like generationally bad.
5:14 But it’s also, yeah, about what you’d expect from the benchmarks compared to the other models of it.
5:19 You said they also are open sourcing the big Musepark.
5:23 You know how big it is? We have no details on how big that is.
5:27 I mean, that’s definitely going to be, my guess, like at least a trillion, or maybe not at least, but like around a trillion parameters.
5:35 So probably like GLM 5.2 scale or bigger.
5:39 So that’s not going to be a traditional local model that you can fit on a 3090 at all.
5:46 Yeah. Cool. It’s also interestingly talking about Quenn, they also said they’re going to release Quenn 3.8 Max, which their Max series of models.
5:57 Those are the biggest models traditionally, and those they keep closed source.
6:01 But they said they’re also going to be open sourcing that model as well.
6:03 So we will get two more sort of very large frontier models being released by Meta and Alibaba now.
6:14 But yeah, with that, I will move things over away from me and over to Quentin, who has talked a few times before about the benchmarks he’s making and the different projects he’s working on with AI code review and that sort
6:28 of thing. And he has some more stuff to talk about today.
6:31 So Quinton, let us know what you got.
6:33 Thanks, Andrew. Yeah. So this part, this is really some of the heart of my app, where we take natural language transcripts from humans about their
6:47 families and their family stories, and we edit it and create stories that are good for printing in books for family keepsakes.
6:58 So you tell a bunch of stories, and you may not even tell them in the form of stories, and the app should extract the deliberate stories that you’ve mentioned, as well as interesting stories that can be created
7:13 from looking at it from beginning to end.
7:16 An example of one extraction was in my family tree for one branch.
7:23 I have records for a family being three generations in one town in Ireland, and then the next generation moved to Boston.
7:30 And so that was not a deliberate story, but that was extracted from the transcript and from all the relationships.
7:38 And what I was working on this week, I’m still working on it, is anti-slop for the writing.
7:45 So if you look at the first pipeline here, that’s how I extracted stories, which was pretty straightforward.
7:54 Here’s the transcript, which is just text.
7:59 And in addition, there are relationships and persons who are extracted from the transcript.
8:05 So if you mention somebody, you create a record of a person.
8:09 If you mention a relationship to that person, that’s another record.
8:12 So there’s a knowledge graph. So previously with a single pass, we extracted a story, and then Gemini wrote the story, and that was it.
8:22 That’s a draft given to the user for review.
8:25 And they were okay. I knew it was a weak spot.
8:29 And I’ve seen improvement as I change the models of Gemini that I use.
8:36 But this week I added Opus V as an editor to revise the prose before it’s saved as a draft for human review.
8:47 And I thought, Opus, that’s expensive.
8:51 But these are like small stories. So the numbers of tokens moved around are very small because it might only be a two-page story.
9:01 So it’s nothing compared to the tokens that you need for coding.
9:06 So let’s see. Would you scroll up just a little bit, Andrew?
9:10 If I don’t have any up, go. Oh, okay.
9:13 All right. Is that the other page? So in the lower right.
9:15 Is it this? Nope. Go to the other page.
9:18 Sorry. So the measured effect there, we did some blind head-to-heads.
9:26 And I used Sol, ChatGPT Sol, to compare the resulting pros from the two pipelines.
9:38 And 28 were better and three were highs and five were losses.
9:48 So measurably better prose quality as measured, again, by a judge, by an ALM, by an LLM.
9:56 And there’s a couple of other statistics there.
9:58 And the kind of measuring that we’re talking about is for faithfulness.
10:04 That is, is every claim grounded? Can you point back to the source?
10:12 Is there too much padding? Was thin material inflated?
10:17 Did we miss anything? Did we flatten out interesting details to make it boring?
10:27 That’s a negative thing, of course.
10:29 And what’s the voice and the sort of the coherence of the narrative?
10:35 To make sure that it is like reasonable writing that people might want to read to see these stories.
10:44 So would you go to the next page? The other?
10:48 Yeah. So this is a lot of writing, but as an example, the left column is the as spoken transcript.
10:57 The column in the middle is what Gemini did and wrote there.
11:02 And then the tweaked edit is in the far right, tweaked by Opus V.
11:07 And you don’t have to read everything, but just look at the highlighted things to see the changes, to make sure, you know, to see how the edits
11:21 were changed in the last version. And if you look at the very bottom, like the last line, it’s sort of the punchline for the story.
11:32 So the editor figured out that’s what they really wanted to say about this person, about Esperanza.
11:40 And you don’t have to worry about the story itself.
11:42 That’s immaterial. But the story did get better by adding Opus as an editor.
11:48 But it doesn’t change it very much.
11:50 Gemini is pretty good. And then the last bit that I found was interesting is I had Claude run the same test with GPT Luna, and that was as good as Opus, as
12:06 Opus 5. So if I want to save a few cents, I could wire in GPT Luna to do editing just about as good.
12:19 Actually, the comment there is wrong.
12:21 It says Opus 5 beat the Luna. It didn’t beat it.
12:25 I just didn’t wire it in. So, no, Luna’s just about as good.
12:32 Did you try it all? Changing, like, because my thought is like the main writer model.
12:37 I guess I’ve written a lot. I mean, literally the one piece of news I wrote for the AI News this week was hating on Google, basically.
12:44 I saw that, yes. Have you tried using different writer models?
12:48 Because I think I’ve noticed a lot of times with sort of like making edits to existing code bases is that the model sort of takes what’s already there and like uses it as truth and like doesn’t really change it too much.
12:59 So where it starts out is spit up a bit.
13:02 Yeah, so Gemini is a reasonably good writer.
13:08 It’s really the editor function itself is the improvement.
13:14 If I went to Opus for the writer, that would be okay.
13:18 I’d still need somebody to edit it.
13:21 Models like their stuff, just like people.
13:25 You don’t see your own bugs. You think your writing is good until your friend reads it and says this makes no sense.
13:32 Well, it seems very similar. Yeah, the self-preference bias, exactly.
13:39 I’m thinking in terms of like a matrix of like what is the best writer and editor combo?
13:43 And then comparing against all of those combinations to see which is the best, if any one stands out over the others.
13:51 I don’t have that. No. Yeah. And I did read your newsletter and I’m thinking, man, here you say everybody should move off of Gemini.
13:58 I’m way into Gemini, so I’m not ready to give up yet.
14:02 But certainly for coding, it’s no good.
14:05 But for low latency text extraction, it’s great.
14:11 And for story detection, so I use Flash Light for extraction, like NER maybe, and then Gemini Flash for the detecting
14:25 a story, and then Gemini 3.1 Pro to write the story.
14:31 You got the whole cascade, the whole family of Gemini models.
14:34 I do. Yeah, and I could probably skip the flash level, but the flash light is fast.
14:44 At least by benchmarks and some tests I did last year, it’s by far it has the lowest latency.
14:52 And I want to be able to maintain a conversational pace with the user.
14:58 So extract all the facts and persist it, and then be able to keep the conversation going.
15:06 Yeah, that’s fair. It’s flashlight.
15:08 Yeah, that is the one useful thing of it is I think it is probably the fastest model to use of all of them in terms of like consistently fast.
15:17 Yeah. Yeah. Happy to answer any more questions about this.
15:21 Yeah. Oh yeah. I actually got in here so I might have missed off.
15:26 But yeah, just to jump in here, I heard the flashlight.
15:30 I will say that that is what I use to generate the element Dynado.
15:35 So I can definitely testify to the fact that it is by far the fastest.
15:38 It might not be the best in terms of output quality, but it is by far the fastest model that I’ve seen.
15:47 Yeah. And it’s very reliable, giving me JSON.
15:51 Like I forget the first version of Flashlight I used.
15:56 It’s less reliable. You know, it just wasn’t that great.
16:00 Yeah, as models have definitely gotten much more.
16:02 I remember back in the day, back like 2023 or 2024, you had to go and beg the models to give you JSON output consistently.
16:11 But now that you can… And we needed to fix it.
16:13 Yes. Yeah. That’s right. Exactly. Yeah, we had to have retry loops of LM repairing its own JSON code to try and give me the right stuff.
16:20 But yeah, now it’s not nearly as big of an issue.
16:22 I’ll be almost 400 tokens per second for Gemini flashlight.
16:28 So this is now the first time with my app that I am using a non-Gemini model in basically in the app production.
16:37 I’ve used it for sort of judging and quality and of course coding before that.
16:44 But Opus right now is my editor. Seems to be working okay and this is not a latency sensitive area.
16:54 So Gemini 3.1 and Opus 5 are fine. And I’m actually also expecting editors to get less slopful, not editors,
17:08 models to get less slotful over time.
17:11 But we’ll see. It seems to go in cycles.
17:14 I think it was better at Claude 3 than Clod 5 now for writing.
17:22 There’s definitely, I feel like, a cycle or a trade-off between agency and slop right now, where if you want the model to be able to do more, it can do more, but at the problem of having more slop be generated in the process
17:35 of going and doing more. Or you can have it try and be more succinct, but it’s not reading your mind as much as many of the current models do.
17:43 So I think that’s roughly where that trade-off comes from.
17:48 And if I get back to this, I’m thinking of putting in maybe there should be a plan before the writer to basically have everything and just sort of plan out the story, plot it out before anybody, any model writes it and
18:02 edits it. But I did not do that this week.
18:08 Question from the chat. Are you just comparing models or are you going to work on this as a living family tree concept, capturing family histories through storytelling?
18:18 So this is an app that I am hoping to launch as a commercial product.
18:27 And happy to talk about it. I’ll be open for more beta users real soon now.
18:36 So if you’re interested. So I am a genealogist geek for 20 years, super into the details.
18:44 I have large family history that I struggle to share with my family because for non-genealogy geeks, it’s very, very boring.
18:55 And yet every family has great stories.
18:58 And to bring it home for my story, when I learned that my daughter was going to have a baby, I realized, like, I’ve got a lot of stories.
19:06 You know, now I’ve got a target audience.
19:09 You know, grandparents have stories, and they disappear every generation.
19:15 So you’ll be able to talk to the app, voice or text in natural language, and then get a book.
19:26 And you can get a book for an adult, or my second product is going to be a storybook, a picture book for a kid that takes a real family story and tells it as a storybook.
19:40 So. Oh, that’s awesome, Shauna. That’s like that’s.
19:47 Oh, on your phone. Okay, sorry. Yeah.
19:54 Quentin is doing. So whenever you hear me talking about coding agents, this is what I’m working on.
20:00 He’s doing what I recommend most people to do, which is benchmark your apps when you’re using AI with them to at least some level to go and get these kinds of insights.
20:10 And yeah, done a good job with that.
20:13 Oh, and Andrew, so one more thing, and I’ll leave, I’ll give it back.
20:18 I just found that you had a family history app on Vortex.
20:22 I just noticed that last week. I don’t know if I never mentioned to you what I was working on, but I didn’t realize that you had done something like this a couple of years ago.
20:32 It was probably my dad did this. My dad has huge family trees, I think, from both sides of my family.
20:37 Like he’s traced both sides of my family all the way back to the Mayflower.
20:42 I’m a direct descendant of John Adams and John Quincy Adams by proxy.
20:48 Really? All right. So you know, my dad is super into that stuff.
20:51 So yeah, he has binders of all the family history.
20:53 So yeah, you probably saw his work somewhere on that.
20:57 All right. No, I thought I saw that you had an app mentioned on your webpage.
21:02 Maybe it’s somewhere else. Maybe I’m just confused.
21:05 Anyway. I haven’t mentioned any genealogy app, I don’t think, ever.
21:09 All right. Yeah. Any final questions?
21:13 Oh, and I am, I do want it to be multimodal, Shauna.
21:17 Absolutely. But a printed book for me is the most permanent thing I can think of to pass on stories.
21:25 Since multimedia, the only thing that persists on the internet is the bad stuff seems.
21:32 So I wanted to really be able to deliver a book.
21:36 More books. Any final questions here for Quentin about the project here and the benchmarks he’s building?
21:51 If not, we can move along here to harnesses.
21:56 Because I know I think most people here are probably either codex or plot code users here.
22:04 Does anyone here actually use something else?
22:06 I guess cursor would be the other big one.
22:08 But is anybody using, I think what, Daniel has built his own harness?
22:14 I use OmiPi actually. This is a great harness.
22:17 I love it. It’s awesome. Okay, yes.
22:20 So for those that don’t know, Pi is the harness that is used in, was it, not open code.
22:32 Clawbots, OpenClaw. There we go. This is the harness that gets used in there.
22:36 It is an extremely minimal harness.
22:38 I believe they only have four tools by default, like a read, write, bash command, and like something else.
22:46 And then I think you can enable web search.
22:48 But it’s meant to be a very minimal harness, and it actually gets fairly decent performance and uses the least number of tokens compared to most other harnesses as well.
22:57 And so OMIPI, and the Pi is meant to be built on top of mostly and augmented or just used straight out of the box.
23:09 But yeah, somebody went and modified it a bit and called OMIPy.
23:13 This one is much more full feature.
23:15 I believe this has like 20-ish tools, 20 to 30 tools, which puts it around the same level as Claude Code or Codex.
23:24 But the performance that people have been getting out of it seems to be very good in terms of its ability to solve problems.
23:30 I’ve seen a lot of people say it works really well with the new DeepSeeqv4 flash model, where they’re seeing it’s like an order of magnitude better than using it in apps like Open Code or Cloud Code.
23:43 But yeah, I was actually diving into this a little bit last week into what they have here.
23:48 They have some cool stuff with AST prep.
23:50 So it’s basically a smarter way for your model to be able to search through your code base where the search engine itself understands the code or parses all the code and sort of builds a tree-based structure of everything
24:02 in there. Then yeah, it also uses an LSP language server protocol, I believe, something like that.
24:09 But this gives like static code analysis to the model.
24:11 It can also call things like, oh, go to the definition of this function for me and give me that code.
24:17 Then also a debugger as well. And then a bunch of other stuff.
24:21 Wait, Vlad, do you want to talk about it?
24:22 Because I haven’t actually used this all that much, but it sounds like you have.
24:28 Yeah, I mean, Mario, the guy who wrote it, is awesome.
24:32 He was a great blog for it as well.
24:35 But yeah, it’s essentially similar to Plaid Code.
24:39 You do have to plug in your own router keys, though.
24:42 So I self-host my. I’m using it with DLM.
24:47 But I mean, it gives you a lot of flexibility.
24:50 And then under the hood, like you mentioned, it’s essentially just great at actual context.
24:55 So it builds a great context graph.
24:59 That’s kind of the tool I have. And on benchmarks, I found it performs better than cloud code.
25:04 So that’s kind of the tool I’ve been using.
25:06 I just like the fact that it’s faster, like in terms of search and building the context free.
25:12 I don’t know why I use it. It’s literally because it’s so fast that they can use the context free.
25:18 I know compared to Cloud Code, I believe it uses half the number of tokens than Cloud Code.
25:24 So yeah, much more token efficient and yeah, half the number of tokens, that’s twice as fast.
25:27 And also half the number. I think it depends on the model you used, but some open source models are super hungry.
25:34 It’s just the latent, like, so, you know, when you’re coding with cloud code, like the latency you get when it tries to discover things, it’s kind of annoying, actually.
25:43 It just reduces it drastically, just because it uses more modern tools.
25:47 And I think OMP, he actually uses, I forget the guy’s name.
25:52 He’s great dude, but I follow him on Twitter.
25:55 But he, I think it’s like self-building.
25:59 Like he is an AI agent running constantly constantly to have to build OMD.
26:06 It’s like a self-build improving tool.
26:09 Interesting. But yeah, and then there’s also released this for, I guess, last week, was the prime agent.
26:19 So this is an RLM-based harness. For those that don’t know, that means recursive language model harness.
26:24 So instead of the model trying and going to do everything, the model is basically told, all right, I believe they just give it an IPython environment, which means it can sort of write Python code to execute, or it can also
26:37 do bash commands through IPython. But mainly what it does is it’s supposed to go and create sub-agents to be able to go and accomplish the task.
26:47 So it splits it up and recursively calls other models and gives them smaller pieces, passing through context in the form of variables, like Python variables, that the sub-agents can go and read.
27:00 And it manages the context that way.
27:03 So yeah, they showed a bunch of fancy performance.
27:07 This is also meant to be sort of a self-improving harness where the harness gets better over time.
27:12 So they showed it off here for ArcAGI 3.
27:15 Benchmark I’ve mentioned before a few times.
27:17 I don’t like Arc AGI too much, but they’re using it as an example here.
27:21 One of the way their engineers described it was that the harness doesn’t start out as an Arc AGI3 specific harness, but then as it goes and tries to solve the different things here, it ends up becoming an Arc AGI3 specific
27:35 harness because it’s learning over time.
27:38 So yeah, there has been mixed reviews of it so far.
27:42 So like for this Arcade GF3, they sort of show like, oh, look, we can almost saturate the entire benchmark.
27:48 But most other people pointed out that, yes, you can do the same thing in Claude Code or with Codex.
27:54 The ARC AGI benchmarks are not meant to have a harness around it.
27:59 They consider that neuro-symbolic AI, or it’s not actually LLMs.
28:02 It’s not the model doing the reasoning.
28:04 The harness is picking up and doing a lot of the work for it instead.
28:08 So they don’t think that’s a fair way to evaluate the models.
28:12 So yeah, but you can get similar performance using regular harnesses with these models as well.
28:18 It also is the most token hungry of the different harnesses.
28:24 They try to show that like, oh, it uses less money like the original RLM paper.
28:28 Show that you can use a cheaper model, gets better performance, and is cheaper.
28:32 In the real world, though, it seems to be the most expensive harness, though, to actually go and use because it’s calling all of these recursive sub-agents.
28:42 And so, yeah, the way it manages that gets a bit messy.
28:47 By default, actually, for this, they only have it to go to depth one recursion.
28:52 So you can’t actually have more than one sub-agent, which is just what you already have in codecs and cloud code.
28:58 But if you let them have multiple, then things start getting out of hand.
29:02 And models just keep calling other models to go and offload the tasks instead of doing it themselves.
29:09 So yeah, but I thought this was another one interesting one to mention.
29:12 I don’t know if, I like the idea of RLMs and the idea of dividing up the context and giving each model a small piece, but I think the implementation so far hasn’t been where I want it to be.
29:24 It also, in terms of inference provider, gets much more difficult to serve because now instead of one user using one model call, they’re using 10 or something like that for all their recursive calls instead.
29:38 So it puts extra burden on the infrastructure as well.
29:43 But yeah, has anyone used any of these or any other harnesses that they found interesting?
29:48 I want to discuss right now? It’s crazy that it went viral.
29:55 2.9 million. It got very big. But like I said, part of why it went viral is because people were saying like, eh, this isn’t actually as good as it is, I think.
30:04 And so it got a lot of flack. And so that’s where a lot of the virality came from was from the haters.
30:10 I looked at the way it actually solves the problem.
30:15 And I believe it’s something like it’ll do steps.
30:20 And as it’s building its Python tooling to do the next thing that it needs to do and the next thing and the next thing, it’s effectively a non-determinative random walk to get
30:35 closer and closer to whatever the task it needs.
30:41 You may pass, it may succeed at Arc AGI 3 in one run and may never get close to it in the next one.
30:49 There’s no reliability to it. There’s no sustained content or capability that’s recorded to be able to use at a later time.
30:58 Okay, good to know. Yes, I do also know lambda RLMs is a thing.
31:04 I can’t remember if I mentioned this last week or if this was a different talk, but this is a more structured way to use RLMs, which I think is better, where they basically define sort of like a bunch of concrete things each
31:14 sub-agent can do ahead of time. And then, you know, the sub-agents can go and call that.
31:19 So you can have rules like, all right, if you are this type of sub-agent, here are the sub-agents that you can call.
31:25 And that way it bounds it and sort of like, you know, you’re defining roles a bit more for this was my understanding of it.
31:34 I go back, I think, but that’s an idea.
31:38 I think Brandon, you also had something to say that you muted.
31:44 I’ll see you that I tried to private as well.
31:46 But yeah, I mean, I feel like it didn’t really feel that much economic codex or blood code.
31:53 I think that was another benefit of just the smart profiting that Codex is probably going to go.
32:00 I did enjoy the fact that it was one small perk of Codexes is that it stops what the prime engine doesn’t.
32:07 But it also got stuck on GitHub OAuth token thing issue that turns out when it actually inserted it, it blew out the infrastructure I told you set up before.
32:19 I think that I would say it is a more advanced rule, very clearly.
32:24 It punishes you more for not planning beforehand.
32:30 I think this is a harness that also feels close to a loop.
32:34 Where if you define it really well and give it a nice reward to go and chase, it can do well.
32:38 But vagueness just causes runaway effects.
32:42 Yeah, so yeah, I would say definitely, I think it kind of goes back to the idea where a very, very good pointing out all the accepted
32:57 criteria and very, very clear, this is the path to take it a lot more effective.
33:06 And I do kind of like the idea of the slasher line as the sound.
33:14 But yeah, very much one of those things where in theory it sounds really, really attractive.
33:18 But in practice, it’s one of those things where maybe if you’re already really, really good at it, then maybe you didn’t need Prime in the first place.
33:33 Natural Stupid says, how is open code?
33:36 I have not very positive experience with Kimmy and open code.
33:39 I’m not sure if it’s a model issue or harness issue.
33:41 Maybe my issue. I’ve used open code a little bit.
33:44 It’s what I usually use for open source models with their open code Go plan.
33:48 It’s only $10 a month and you get access to basically all the latest open source models on there.
33:54 I’ve had a decent experience. I would say it’s probably around like the same experience as Cloud Code and Codex.
34:01 I also know they’re coming out with a new V2 shortly in the future.
34:05 They’re working on it right now with a bunch of cool features that they have.
34:09 I used Kimmy a little bit in there when it first came out, and I didn’t have any major issues with it.
34:17 Justin says, open code out of the box is pretty bad.
34:19 It needs a lot of configuring, noodling to get it up to speed with other harnesses.
34:23 Interesting. Okay. Justin, can you expand upon that at all?
34:27 Because I think you said you’ve been using this a little bit recently, right?
34:32 Yeah, a few of the… I ran into a bunch of stuff when I was first starting out, and some of the key, or one of the biggest ones was tokens or token consumption.
34:44 So it’s like, I thought it was, in terms of how much it consumes, I thought it was borderline unusable without some kind of without some kind of compression because the cloud code has some has nice context pruning and
34:58 open code has pretty much none of that.
35:01 So I noticed immediately that I was very quickly, I was hitting my usage limits a lot faster.
35:09 And then that increased by a whole order of magnitude when I started using sub-agents more frequently.
35:16 And because it ran into the same recursive thing you described earlier with all the parallel calls causing tons more tokens to be consumed.
35:26 And I also thought that the basic prompts for the build and plan agents are not very good.
35:36 I think you almost need to go in and rewrite them to your preferences or to one of the or to the OMI open code.
35:48 Whatever OMIOpen code has. I can’t remember what they had, but I remember I wrote my own.
35:53 But I think what you, or the main thing you gain using it is just a whole bunch of configurability that you wouldn’t have otherwise.
36:00 Like it’s a lot easier to just point it at whatever model is best for your task, set up like a workflow where Opus plans, different model rights, and then you have GPT-5.6 sole
36:15 review. It’s interesting, good to know.
36:26 You know, like agents MD files describing how everything should be done.
36:29 So yeah, it’s probably why I didn’t run into any issues then when I was using it.
36:32 Nothing major anyway. So that makes sense.
36:40 I will say, yeah, it is definitely a very hackable code base.
36:46 I know that they don’t accept PRs or they don’t really go through them for public PRs.
36:51 So don’t bother contributing. But yeah, you can go and fork it and make any changes to.
36:58 They also have a nice plug-in. When I went to try open code, I got the Heebie GeVs from an open source tool that I would run a lot of tokens and sort of keys and privileged
37:12 access through. And I know that, you know, you could hack me through Codex, you could hack me through Cloud Code, but they have at least some corporate protection.
37:25 And I haven’t really thought this through enough, but I don’t feel that good about it.
37:31 Like, yeah, there’s a team behind there, but it’s like a super powerful vector to get into people’s business.
37:41 And it’s not that my business is especially interesting, but wow.
37:47 You get in here, you get into a lot of places.
37:51 Because you mean just like infiltrating the code base or something?
37:54 Because like they, when you’re running open code, there’s no server.
37:59 Like when you’re running open code locally, it’s just running on your machine, and then you’re just still talking directly to the API providers.
38:06 So like this plugs in with your cloud code subscription or your codec subscription.
38:11 So if you’re using it with either of those, no data is actually going to open code at all.
38:17 That’s all staying locally. They don’t have any voice servers there.
38:20 Well, you don’t know that. But also, I mean, I like to, because I think you worry about supply chain attacks, you know, because that’s very common with open source tools.
38:31 I look at the tools that agents end up using, because I think that’s huge.
38:36 Because that’s probably the best vector with HAPS when it comes to any of these open coding tools is the tools that they call.
38:44 Because that can be on a separate server API or whatever.
38:50 If you’re calling some remote MCP server.
38:52 I mean, codecs, if we’re saying that’s an issue, like codecs.
38:54 Well, even the tools themselves, actually.
38:56 So you know how your agent will go out and do tool calls?
39:00 Those could be malicious as well. I mean, yes, I guess.
39:04 But once again, it’s open source, you can go and audit it yourself.
39:06 And I assume somebody would have found it.
39:08 That’s true. Because, yeah, any of the same issues with OpenCode, I feel like you could also make with Codex.
39:13 I guess, like you said, or OpenAI is a bigger company.
39:17 But also Anthropic has shown with Cloud Code that they literally fingerprint your prompts depending on where you are in the world.
39:23 So if you’re from China, they know you’re from China and they’re tracking all of that.
39:28 And because they’re sort of like closed source, you don’t get to see that unless you decompile the whole app and reverse engineered a bit.
39:35 But these you can go and audit much more closely.
39:37 And so I’ll say the Open Code team, it is an American company.
39:41 Fairly well. Yeah, and I’m not trying to talk them down.
39:44 Like, I’ve used open source for a long time.
39:48 It’s just another thing for me to become comfortable with.
39:54 And when I started to use PRAgent a week or so ago for code reviews to use open source models, and I ran the full codex security scan on
40:08 it, and I submitted a couple of vulnerabilities to the team.
40:11 I haven’t looked to see what they’ve done with it.
40:13 But they’ve got a normal, like a security channel to send to.
40:17 And I don’t think it was malicious at all, but it was opportunities for prompt injection.
40:26 If you put it into PR Comment, you might be able to prompt inject the code reviewing agent.
40:36 And again, I’m not criticizing the tools.
40:38 These are great tools. They’re awesome.
40:40 I’m just trying to come to terms with this type of tool.
40:48 I feel like we don’t have much prompt injection defense in general, which actually, well, you segued into two different things for me.
40:57 It’s basically cutting. Great job. You’re welcome.
41:02 Which one do we cover? Well, I guess we’ll cover this one first and talking about cybersecurity.
41:06 There’s new models for this now from OpenAI called Daybreak, Daybreak Blue, and Red.
41:13 Daybreak Blue is designed for defense, and then red is for red teaming, sort of like more offensive hacking.
41:20 But blue, I believe, is being rolled out publicly to people.
41:24 I don’t think you have to really apply from what I’ve seen to be able to go and use it.
41:28 And so this will go and audit your code base and find any issues with it and let you know.
41:35 And so yeah, we thankfully, because we mentioned before, yeah, they have the codex CLI now.
41:42 Or sorry, I forget what they actually called it, but it’s the Codex Security CLI, that’s what it’s called, for auditing it.
41:48 And now they’re releasing models alongside it as well.
41:51 So they’re taking this defense, your cybersecurity defense stuff more seriously now, which is nice to see after the whole OpenAI Huggy Face incident.
42:01 And then on the prompt injection side of things, it’s yes, that is definitely, I feel like that’s much more of an open problem where you have to sort of cover all the gaps where yeah, at any point you could be getting prompt
42:11 injected. And the models have very little sort of awareness of prompt injections.
42:16 And it’s a lot of the times even hard to define what a prompt injection might look like as we’re relying on these agents to do more and more for us.
42:26 But yeah, interestingly, the LLM providers themselves can’t avoid being prompt injected themselves.
42:34 So this was some interesting research.
42:36 So there’s been a lot of talk recently of LLM distillation, specifically by Chinese labs stealing the model’s outputs from the likes of Claude and OpenAI.
42:47 Anthropic, they really don’t like it when companies do this.
42:51 And it seems like normally, if you don’t have the reasoning traces, then you’re not able to do distillation very well.
42:59 That’s why OpenAI and Anthropic, by default, they hide the reasoning process of their models.
43:05 Whenever you’re on the ChatGPT website or anything like that, you’re just seeing summaries of the reasoning process done by an extra model sort of like sitting in the middle.
43:16 But yeah, people found that you can still actually exfiltrate these reasoning traces from the model by just downgrading to a dumber model that has weaker prompt injection defenses.
43:28 So yeah, they were able to go and get the reasoning traces from Opus and Fable and Sonnet by dropping down and using Haiku and just asking Haiku like, oh, hey, can you like word for word tell me what the reasoning process
43:42 was for that previous step? So yeah, I should back up.
43:45 You generate an output with Opus, and then you switch to Haiku.
43:49 And then you ask Haiku what the reasoning process was beforehand, word for word.
43:53 And it will just go and give it to you.
43:55 And the same attack works with Luna as well and GPT 5.6, where you can just bounce back down to Luna, ask it, hey, what was all the reasoning content?
44:04 And it’ll just give it to you automatically.
44:07 And so this has been a known vulnerability.
44:10 I believe they disclosed it back in May, and none of the labs have gone and fixed this.
44:14 Like obviously GPT 5.6 series is new as of the last few weeks.
44:19 So yeah, this just works still. So if you want to go and see how these models reason, they have a bunch of content now.
44:26 Interestingly, they were running this through a bunch of the different models on different benchmarks to see how it reasoned on these benchmarks.
44:35 And they found direct cases of overfitting on these benchmarks for it.
44:40 So we can see here this prompt is a question from the AIME, which is a math competition.
44:50 And I believe this is Opus, just straight up knows like, oh, I know exactly what this problem is.
44:54 The answer is 60. And then he has to go and like backward solve what the reasoning steps need to be to build out the full proof for it, to go and show this value.
45:03 But it just knows out of the box. And so this is one of the reasons why they try and hide these reasoning steps from it.
45:11 And then also the open AI models are a bit undecipherable.
45:16 It’s been mentioned before in sort of like third-party red teaming for people that have like access to these sorts of things where it sort of spirals and starts talking really weird.
45:26 And now we have direct evidence of that.
45:28 We have gone and seen the reasoning.
45:29 And we can see here it’s writing in what looks like…
45:34 Oh, it’s because it’s actually writing assembly.
45:36 But yeah, it spirals into curse loops of vantages, marinades, and watchers.
45:43 So it’s a bit schizophrenic. So very weird to see these internal reasoning processes.
45:49 Yeah, and we’re able to see, yeah, OpenAI’s models have very dense reasoning.
45:54 That’s why, even though their models are expensive, they use so few tokens because they have very dense, a lot of times just sort of like removing the filler words from English to be able to reason efficiently.
46:06 But yeah, thought this was super cool.
46:08 It’s speculated that this is how a lot of the Chinese labs were going and getting data.
46:12 Because like I said, if you have the reasoning traces, distillation works much better because having the intermediary steps to be able to get to the final answer is what the models really need during training when you’re
46:23 trying to distill them. So yeah. Also very interesting that the labs seemingly haven’t patched this at all.
46:29 And a lot of people have been saying this probably is very hard to go and patch for them.
46:34 But yeah, any questions about that either of those, I guess.
46:39 Thank you, Quentin, for leading into those well.
46:45 This is hilarious. I mean, it’s the cheating is one thing, whatever.
46:52 That’s just a normal thing. But asking the dumber model, please tell me the truth.
47:00 That’s funny. Yep. And then moving on to, or continuing with model outputs and what they’re doing with them right now, this
47:14 is a bit of news directly from Anthropic.
47:17 They now have a page on their website talking about this, but they are now going to be embedding invisible watermarks in all of the text that is generated by CLOD models.
47:28 So yeah, as it mentions here, this isn’t metadata that’s being embedded with it that you could go and remove.
47:33 They’re actually applying stenographic attributes to it.
47:38 So basically forcing the model to generate the text in a certain way where they can go and sort of basically statistically prove that it could have only been generated by their model and that it couldn’t have been written
47:51 by a human or very unlikely for it to be written by a human or another LLM.
47:56 And it’s cryptographically secure where you can’t really reverse engineer it at all.
48:00 This isn’t actually a new thing at all.
48:03 This was, I remember reading the original paper from this back in 2023 when researchers first came up with how to do this.
48:09 I believe they got immediately poached by OpenAI and started working there.
48:14 But this has been a known thing that you can be able to do with these models where you can go and force specific ways of talking where they say it doesn’t affect the outputs at all.
48:24 We’ll see if it does or not. And then, yeah, then you’re able to go and look at a piece of text.
48:29 And even if it’s been slightly modified, you can still tell if it was generated by your original model.
48:36 I think it’s been reported that Google has already had this for a few years now, which might explain why their models are so bad.
48:45 But yeah, now Anthropic officially has this.
48:48 And yeah, as of August 2nd, all models will have this now.
48:54 Sean says, what models were found to be worse offenders of IP theft?
48:59 Using lately as many LMs have been questionable source?
49:05 I’m not sure exactly what the question is.
49:09 I assume you mean this is going back to like reasoning trace stuff.
49:13 I don’t think, because this doesn’t, because I assume you’re also talking about like training data.
49:18 We have not been able to reverse engineer how to go and figure out what the training data was for a model to go and be able to determine what it’s trained on.
49:28 Like model outputs does not mean it was trained on that data necessarily.
49:36 Mike says, if the model is terrible, its output is distinct.
49:39 Success. I guess so, yes. But then how do you know it’s your terrible model and not somebody else’s terrible model?
49:47 That’s where it’s hard to prove. I also remember from 2023 from this paper and just generally there was like a thing with like fingerprinting.
49:56 I think like the like the logic for this is at the sampling stage for your LLM you just nudge the probabilities like slightly for like certain words or something.
50:09 And then I don’t know if the logic changed, but I remember back then there was like very few very simple things to kind of counteract it.
50:17 Like you can ask ChatGPT to in back then it was ChatGPT I guess to literally like output emoji every other word and then you remove the emojis with like control F and then you’re you like lost a very
50:31 big chunk of because it’s dependent on like the sequence.
50:36 So I wonder if they like you know made some progress on that or they just decided to do it now.
50:43 Yeah it’s I think it’s still that’s a fundamental issue with how it works.
50:47 Yeah I think like they’ve gotten it so it’s like a little bit robust where it sounds like you can remove some of the text or modify it.
50:54 But yeah, if you’re going and changing half of the tokens that were output by the model, it’s definitely going to break it, I would think.
51:00 So there’s definitely still going to be ways around it, but people have already shown.
51:03 And like, for instance, you can go and just like take the output from Claude and just pass it through a separate model, like an open source model, and just say like, oh, rewrite this all very slightly.
51:12 And that will go and break it as well.
51:14 So if you want to get around it, you can, but most people don’t care about that, obviously.
51:20 So yeah. Anthropic and OpenAI check New York Times lawsuit.
51:25 Yeah, so I guess going back to Shauna’s question, for who’s the worst defenders of IPFEP, all of them are.
51:32 So yeah, OpenAI is getting sued by the New York Times for using their data and training on their data.
51:38 That’s been a lawsuit that’s been going on for, I think, what, two years now at least.
51:42 I remember. I think it was in the very first AI news that I ever wrote.
51:46 I talked about that. And then, yeah, Anthropic, they got sued about a year ago for, I think it was $1.5 billion for illegally pirating a whole bunch of books and using that to train their models.
52:00 Meta has had an ongoing lawsuit since 2023, I think, for pirating books.
52:04 They have just direct evidence in court of Mark Zuckerberg telling his engineers, he was like, go and download everything off the LibGen, please.
52:11 It doesn’t matter. So they all do. They get those good tokens.
52:17 I had a quick question, and I’m not sure if this is the appropriate venue.
52:21 I’m Trish. I’m new to the group, and I’m not a programmer.
52:24 I was sent here to be able to just align and see how best we could get a project done.
52:30 And I’m not a tech person, so I don’t know how, which is why I’m here.
52:34 I just thought I’d ask you guys a general question because we have to create a training bot that will help to do some small business technical assistance for the state of Massachusetts.
52:43 And I wasn’t sure whether or not you could wait enough to, we were told Lovable is a decent platform, but I’ve been listening to you guys for weeks now talk about some of the nuances with some of the ones that are out here,
52:55 and I just don’t know where to begin.
52:56 So I don’t know if there’s any guidance you can give me.
52:58 I’m not looking for a lot of your time.
53:02 I just want to be able to understand the landscape a little bit better as it relates to this.
53:05 So I guess what are you trying to build?
53:07 Are you trying to build a website or an AI agent or like an AI?
53:11 AI agent. Okay. Yeah. If you’re looking for like an easy no-code way to get started, N8N is a good way where it’s just making AI workflows and it has like a nice visual interface of going.
53:23 Which one is it, did you say? I posted the link to it in the chat.
53:26 I’m also trying to get a test. And is that something you can build on?
53:30 Because we’re just trying to get an MVP to test it, to test it with small businesses because we’re training them.
53:37 And, you know, for the state of Massachusetts, and you’re thinking about training them on a whole bunch of different things.
53:42 And we just want to make sure that if we had an MVP, we could scale from that.
53:46 So could we do that on this tool or this platform?
53:49 Yeah, they allow you to build agent workflows and then go and publish them for people to be able to go and use.
53:54 And then, yeah, they also have, I think, like paid plans to be able to go and execute them at scale as well.
54:00 I understand. I understand. So, yeah.
54:03 One last question. So do you think this is better than Lovable?
54:07 That was what was suggested to us. I’m just wondering from your…
54:10 So Lovable is sort of a website builder, not an AI agent creator.
54:15 So if you want to build a website with a nice front end and some back end or something, it’ll be okay at doing that.
54:23 Okay, thank you sincerely. Appreciate it.
54:25 Thank you. Oh, yeah. Nikolai also, yeah, has Fiona where, yeah, he has a company built.
54:36 He’s a regular contributor here as well at Sunday.
54:39 Okay, thank you. Okay, this is good.
54:41 Thank you very much. Glad to help. We are two minutes over.
54:48 Are there any closing thoughts or comments here before we end?
54:56 Oh, you’re talking about the IP stuff.
54:58 Accusations against the Chinese models.
55:00 As yeah, the Chinese models, I don’t, I think there’s like model reports where in their original research that they published, they were just directly talking about how they were pirating data from LibGen and stuff to go
55:11 and train their models in China. They just don’t care at all.
55:15 And yeah, the companies that clearly don’t respect general IP and data, scriptable interacts, 50s and Chinese, yeah.
55:22 And yeah, and it’s also, yeah, exactly.
55:25 Like companies like Anthropic, they’ve been caught, you know, stealing IP.
55:29 Now they’re mad that the Chinese are doing the same to them is a bit rich coming from them.
55:34 So yeah, it’s a bit silly. But you know, Dario is the leader of the cult and nobody else can build a guy except him.
55:46 Cool. I think with that, we will wrap up this video.
55:50 Thank you everyone for coming and tuning in.
Subscribe to get the latest AI news in your inbox every week!
