Peter @petmouse.bluesky.mousses.xyz · 7hgo to your camera roll amd search "boat" and post one that you like 020 2
Reposted by PeterGus @gusthema.bsky.social · 17/10/2026What would you like to see on Gemma 5? 24685 68
Peter @petmouse.bluesky.mousses.xyz · 11hnot sure if this is already done, but having experts dedicated to verifying tool calls might boost accuracy there native audio input on MoE models as well as native audio output. local assistants would feel so much smarter and faster if they didnt have to use a separate emotionless TTS system 2/2 050 5
Peter @petmouse.bluesky.mousses.xyz · 11hmy main pain point with the Gemma 4 family has been malformed tool calls. having some MoE experts that improve this would be awesome for small assistant models on the topic of MoE, it would also be great to see something like 50b-a4b squeezed into edge hardware with SSD expert streaming 1/2 270 7
Peter @petmouse.bluesky.mousses.xyz · 18/10/2026i was talking more about the session token. analogy would be each key is a session, then when they need a new one it can be provisioned after auth (like provisioning a new api key in an admin ui) 010 1
Peter @petmouse.bluesky.mousses.xyz · 18/10/2026i tried to one-shot some architecture diagrams the other day, and idk if this was just a skill issue. astra was the only one to get close, and even then i had to iterate to get a very basic diagram that was at least fundamentally correct. even then i still needed to make my own 000 0
Peter @petmouse.bluesky.mousses.xyz · 18/10/2026quoting because I can't comment. isn't this similar to logging out, just with an API instead of a UI? 110 1
Peter @petmouse.bluesky.mousses.xyz · 18/10/2026obviously this doesn't sustain when most jobs are automated, but I'm sure a lot of people who do got shuffled around inside a company are in a UBI-like job, aka a relatively useless powerpoint job. I'm interested to see how it plays out 000 0
Peter @petmouse.bluesky.mousses.xyz · 18/10/2026AI itself is inherently anti rent seeking. I don't know exactly how it will work when every job is automated, but if we don't start early we don't have runway to fail China has an interesting policy for this that makes it illegal to fire someone because their job was automated ... 100 1
Peter @petmouse.bluesky.mousses.xyz · 18/10/2026UBI is a really important step before we start mass automating jobs only do e for the paycheck. the only alternative is societal breakdown the countries that get UBI right will leapfrog everyone else 1123 12
Reposted by PeterAstra ⎔ @astrra.space · 17/10/2026huggingface.co/deepseek-ai/... happy deepseek day to anyone celebrating 716015 160
Reposted by PeterJoshua White @jrw14.whnc.me · 17/10/2026New day. New DeepSeek. Introducing DeepSeek V4.1 Flash 1017. At this point, what is the argument for NOT being a fan of this lab? Full MIT-licensed shadow-drop of models like this just insanely awesome. HF: huggingface.co/deepseek-ai/... Update: api-docs.deepseek.com/updates/#dee... 315619 156
Peter @petmouse.bluesky.mousses.xyz · 17/10/2026this is why local abliterated models are the future 010 1
Peter @petmouse.bluesky.mousses.xyz · 17/10/2026I disagree. Tesla's innovations back in the day, like the octovalve, actually pushed the range as far as it could go. now others have caught up. Tesla recently refreshed the Y, but didnt put in an 800v architecture, which all the competitors have. theyve all caught up atp 010 1
Peter @petmouse.bluesky.mousses.xyz · 14/10/2026if you think hard enough, a pds is like a bucket 140 4
Peter @petmouse.bluesky.mousses.xyz · 13/10/2026google glasses mod to show codex usage limits in the corner like health and mana send tweet 140 4
Peter @petmouse.bluesky.mousses.xyz · 13/10/2026type of guy who thinks collecting the training data is AI's original sin 050 5
Peter @petmouse.bluesky.mousses.xyz · 12/10/2026not sure why this is a better strategy than just prosecuting them for crimes committed on the news that already have extensive coverage 050 5
Reposted by PeterEric Umansky @ericumansky.bsky.social · 11/10/2026This is exactly what ICE said: "Parra-Hernandez ran in front of an ICE vehicle and was struck." Now watch the video. They're no other word for it: They're lying. 16741211791 4.1K
Peter @petmouse.bluesky.mousses.xyz · 12/10/2026the load-bearing assumption is that they share the same substrate. Frito-Lay is public, OpenAI wishes they were 000 0
Reposted by Peterjeffery --dangerously-skip-permissions @jefferyharrell.bsky.social · 11/10/2026AI image generation is one of the good examples, I think. This tech is *everywhere* now, but you don't see it when it's used well. I'm not talking about image generation per se, but patching and in-painting, stuff that used to have to be done by hand. Tremendous innovation. 3202 20
Peter @petmouse.bluesky.mousses.xyz · 10/10/2026yeah, its just that a lot of people haven't actually automated it. it definitely is possible though 040 4
Peter @petmouse.bluesky.mousses.xyz · 10/10/2026fun fact, most banks require telephony MFA also fun fact, phone companies only require a 4 digit code to transfer a sim 040 4
Peter @petmouse.bluesky.mousses.xyz · 10/10/2026m.youtube.com/watch?v=dK5f...m.youtube.comULTRON's AwakeningYouTube video by Soulja22 000 0
Peter @petmouse.bluesky.mousses.xyz · 10/10/2026called it as soon as i saw he has a specific word count goal for his newsletter 010 1
Reposted by Petermr. TIM @timkellogg.me · 09/10/2026whoah, congrats @bcantrill.bsky.social imo Oxide is one of the coolest non-AI companies out there. Great to see 0534 53
Peter @petmouse.bluesky.mousses.xyz · 09/10/2026i love codex's desktop app but it seems to be heading in a bad direction whatever prompting they have in there makes the models constantly try to access UI apps, even when a better cli is available. saw this w/ docker :/ also why is plan mode hidden and why do i need to re-bind the shortcut for it 010 1
Reposted by Peteraly @aly.codes · 09/10/2026is is extremely hard to effectively follow AI right now without at least following accounts on Bluesky that report tweets from X by a couple people. there’s a lot of reasons for this, but it’s true. 4682 68
Reposted by PeterMatthew Sanabria @matthewsanabria.com · 09/10/2026We raised a few bucks. We'll try not to spend it all on DRAM. Thank you to our investors for this round and thank you to my amazing colleagues for building a great company worth investing in! oxide.computer/blog/our-445...oxide.computerOur $445M Series D | Oxide Computer CompanyOur Series D round of financing. 1895 89
Reposted by PeterOxide Computer Company @oxide.computer · 09/10/2026Today we're announcing our $445M Series D and announcing that we've reached profitability. Huge thanks to Eclipse for leading the round, which included participation from existing investors USIT, Jane Street, Riot Ventures, Intel Capital, Friends and Family Capital, and Counterpart Ventures. 727921 279
Reposted by PeterElena Rossini ⁂ @elenarossini.com · 09/10/2026Dear friends who don't speak German: Go to YouTube ASAP, turn on subtitles → autogenerated translation and watch this video! TL;DR: #WSocial is posting legal names, nationalities and approx. birthdays of verified users as a public record on the Atmosphere. Confronted, they claim it's "by design" 🤯 1019991 199
Reposted by PeterJauntyWunderKind @jauntywunderkind.bsky.social · 08/10/2026having a big ai model write a ton of big plans, then forking the session and using a (perhaps) small model to chat about it, learn about it, dissect it is so unbeatable. dont poison the context, chat out of band, learn. out of the main context. with a perhaps not so dastardly expensive model. 1273 27
Peter @petmouse.bluesky.mousses.xyz · 08/10/2026replace the word "LLM" with "5 orbs" to check if your use case is insane ✅ I use 5 orbs to analyze meeting notes ❌ 5 orbs read bedtime stories to my kid 140 4
Peter @petmouse.bluesky.mousses.xyz · 08/10/2026fun fact, anyone with 250k to spend on a property in Greece is eligible for citizenship 000 0
Reposted by PeterDoll @dollspace.gay · 08/10/2026As always the bad guys already have accelerated offence while we have hands tied behind our backs defense 4343 34
Reposted by PeterLeah McElrath @leahmcelrath.bsky.social · 08/10/2026Large banks in South Korea were hit by a cyberattack last week. The hacker appears to have been a single individual—an amateur—who used an open-source Al penetration tool named ARTEX, DeepSeek v4.1-Flash, GLM-5.3, Grok 4.6, and Claude Code. 311945 119
Peter @petmouse.bluesky.mousses.xyz · 08/10/2026same thing except framework desktop vs $200/mo sub 050 5
Peter @petmouse.bluesky.mousses.xyz · 07/10/2026tried it for a bit, then stopped when i realized getting pulled over would mean handing my unlocked phone to a cop 010 1
Peter @petmouse.bluesky.mousses.xyz · 07/10/2026evergreenm.youtube.comLouis C.K. Defends MasturbationYouTube video by deejayPstar 000 0
Reposted by Petermax power 🌌 @sneptech.bsky.social · 07/10/2026the biggest AI misalignment risk is pointedly not anthropomorphising them wherever convenient 0937 93
Reposted by PeterDoll @dollspace.gay · 07/10/2026If you think the courts can save your shitty product people can vibe fork it faster than you can issue takedowns 0254 25
Reposted by PeterNathan Lambert @natolambert.bsky.social · 06/10/2026Too many people are analyzing the cyber risks of open models in a narrow lens, assuming China doesn't care about safety, and fear mongering about private information on open-weight cyber attacks. The public information we have paints a very different picture. www.interconnects.ai/p/the-cyber-...interconnects.aiThe Cyber Risk Discourse is BrokenOpen-weights, ideology, and acknowledging trade-offs. 0235 23
Reposted by PeterJake Gold @jacob.gold · 06/10/2026Ignoring that it's DHH, this is really just a typical example of how many developers are generating code now. Sometimes it's totally fine and other times its kinda scary. Always a huge advantage for the rest of us who continue to maintain a strong understanding and control over our generated code. 4512 51
![Andrew Curran @AndrewCurran_
X.com
Last week some of South Korea's biggest banks were hit by a cyberattack. Thanks to a report from CrowdStrike tonight, we now know the entire hack may have been done by a single person. He used a combined stack of an open-source Al penetration tool named ARTEX, DeepSeek v4.1-Flash, GLM-5.3, Grok 4.6, and Claude Code.
The Claude Code sessions reveal a two-server architecture: The Hong Kong-based IP address serves as the primary attacker-controlled infrastructure, and the IP address 38. 244.50 [• ]120 hosts the ARTEX instance likely responsible for the described Korean attacks. The ARTEX instance used DeepSeek v4.1-flash as the primary LLM backend, and the threat actor supplemented this LLM with GLM-5.3 (Zhipu Al) and Grok 4.6 for additional Claude Code sessions. The threat actor likely accessed DeepSeek via the likely LLM API proxy/reseller xcai[• ]pro.
3:14 AM • 10/8/26 • 563K Views](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:gvda6fem6r7selm4gzjjww4a/bafkreic2ecy4knref6xt6tfc27brgigk6e4pj7lx6vnmeduhh23dnkjnum)
