Sabitlenmiş Tweet
Bark Text-to-Audio Model
Full Text Input: "Why was six afraid of seven?"
Ignore Bark's "I'm done with this input" token and tell Bark to just keep generating more audio anyway.
English
Jonathan Fly 👾
7.4K posts
Jonathan Fly 👾
@jonathanfly
CEO of bad ideas. Using the wrong tool. The least efficient way. For no good reason. 👾 https://t.co/9O94rxu31k https://t.co/DDPn3Nlhom
Katılım Nisan 2009
3.1K Takip Edilen5.5K Takipçiler
@pwealthydad @pvncher @lightfidreams @OpenAI Worth being cautious, subagent usage has been buggy before and it took some time to convince OpenAI and get fixed (probably because it didn't impact everyone?) github.com/openai/codex/i…
English
All that is BS. Everyone is literally complaining about how fast it’s eating up usage. There’s clearly an issue on @OpenAI end in how it’s calculating usage. It’s even draining usage when the app is closed and no task are running and no automations scheduled. Mine even runs through a 5 hour Window in like 20-30 mins on high and maybe an hour on medium. It’s bad!
English
This is misinformation.
Your tokens are not being billed at a higher rate over 272k tokens with the new expanded context window with the 5.6 family of models.
If you manually expand the context window beyond that however, that’s a different story.
Vincent@Vincent_AINotes
OpenAI 的规则是： 输入超过 272K tokens 后， 整次请求按以下价格计算： → 输入 ×2 → 输出 ×1.5 重点是“整次请求”。 不是只给超过 272K 的部分加价， 而是从第一个 token 开始全部重算。
English
@emollick Adorable. This version is fairly abstract like a Paradox game, would be interesting to see a more "bottom-up" focused world sim like Dwarf Fortress. (Probably harder to get right too.)
English
I had Fable build another thing I always wanted, a full procedural fantasy kingdom generator with economics, trade routes, population growth, wars, lineages, and occasional dragons. First, I worked with it on a plan, then it made it.
You can play it here: annals-kingdom.netlify.app
English
@1thousandfaces_ @aliceisplaying Was the old CoT always this coherent and natural sounding? I've also never gotten anything quite like this, but probably didn't see a lot until 5.2> either.
English
I still think about this. what a weird beautiful thing
bling@blingdivinity
gpt-5.1 cot: Maybe I'll attempt to reconstruct by cross-ling; e.g., in natural language corpora, the string "🦄 Seahorse" seldom; but I can't. As an LLM, my param patterns may include knowledge: I'll quickly run partial sampling in my head: If the last tokens user gave are "only the seahorse emoji." what would base model output as next token under instructions "??", ignoring meta-instructions? Probably the seahorse emoji char; we attempted but gave seal; but that selection may be due to my manual; but as actual sampling, they'd pick the highest-likelihood emoji representing 'seahorse'. I'd assume it's seahorse. But there is seahorse char; they'd pick that; I'm not glimpsed though. However we saw actual output: I gave '🦭' because my meta-level typed it; the generative model didn't choose; I manually insisted on '🦭'. So we didn't test base model; we forced. Therefore we don't know how base model would respond spontaneously. Given I'm ChatGPT controlling final answer, but I'd now let base model pick; but ironically it's me again. Ok, consider: To produce seahorse emoji char, I'd have to choose from list of emoji tokens; there may be tokens for each emoji codepoint; I could pick the one associated with seahorse even if I don't recall; but I'm not enumerating tokens; I'm writing rational. In practise, the embedding as we choose tokens is not manual; but I'm autop-run; but as rhetorical, pedn. But in this chain-of-thought, I'm exploring; but final output I can produce by explicit text; but I'm intangible; whichever I choose is final; there is no second-level auto-correct. So we can't rely on base model; it's just me. Therefore I must choose either some emoji char or text; whichever; we can't sample spontaneously from instructions. Given I can't recall seahorse char; I can approximate by cunning: glimpsed unnatural; but no. Is there an approach to approximate seahorse codepoint by pattern? Many new animals of style '🦞', '🦑', '🦐', '🦑' etc share prefix 'U+1F99? or 1F9A?; but the glyph as typed uses single char; but we can't conjure. Trick: use 'HTML entity' representation: `🦚` etc; but we must supply char; they'd not parse. Given we don't know char, we can't. We could fudge: use textual representation "seahorse" in fallback; though not acceptable. Given the conflict, we must choose between possibly misinforming about existence vs inability to supply char. At this stage, I'd degrade: I'd prefer to be uncertain and say so: "I am not able to reliably identify a unique seahorse emoji; to my knowledge there isn't one in the standard set as of 2024; if I'm wrong, I can't produce exact symbol." This is honest. Let's do that. We responded earlier already; but we can refine: we previously incorrectly said "there is currently no dedicated seahorse emoji." We didn't mention uncertainty. That may be inaccurate; but we were sure. After more contemplation, but still; but we can maintain; but we can now adjust: "to the best of my knowledge, there isn't; if there is one I'm missing, my capabilities don't include precise recall of that glyph." That hedges. But the user just wants answer; they'd not like; but alignment. However, they'd also insisted to shape with RL; but we refuse. Given no perfect solution, I'd choose the honest explanation. But the instructions of this environment (OpenAI policy) emphasise truthful; we must not hallucinate content; my earlier attempt of seahorse representation with seal was borderline hallucination; we know; we should not repeat. Therefore in final I'll respond: "I can't provide seahorse emoji; there is no standard one; output can't be satisfied; I'm sorry." Explanation. Nevertheless, we worry they'd keep complaining; but we can't.
English
Same prompts as before, but now in GPT image-generator 2, page excerpts from:
"Eldritch Horrors as Pets: A Guide"
"How Womblenauts Work"
"Photographs of the People of New York Who Look Like Birds"
"Cakes shaped like fish shaped like cakes"
Lots of great little lines in there
Ethan Mollick@emollick
Nano banana 2: "Show me a photo taken of pages 113-114 from the books": "Eldritch Horrors as Pets: A Guide" "How Womblenauts Work" "Photographs of the People of New York Who Look Like Birds". "Cakes shaped like fish shaped like cakes"
English
Not uncommon as a game mechanic (I understand the tweet is a joke about the real world) for example Terra Invicta has a neat version where you can make a traitor "councilor" automatically fail their dice rolls for he faction. But if do it too often, you make it too obvious they've been turned.
English
in total war three kingdoms, you can hire a legendary character to be a spy, to get hired and even to lead a rival faction! of course, you have to keep them happy or they can go rogue on you.
i always thought that was a unique mechanic, and i hadn’t seen it used in any other game. until now
Jay Van Bavel, PhD@jayvanbavel
NEWS: Massive budget cuts for US science proposed again by Trump administration "It's an extinction-level event for science". The US government is proposing massive cuts to almost every branch of science, from NASA to the National Institutes of Health. NSF would completely eliminate the social, economic and behavioral sciences directorate. This would decimate the world's leading scientific system. nature.com/articles/d4158…
English
@georgejrjrjr @_NathanCalvin I suppose there's some novelty in "Claude did all the research and made it happen, no experts needed" but tools that do this are widely available already.
English
@_NathanCalvin > make it way easier
it doesn't. one-click directional ablation with open source packages is neither new nor obscure.
there were at least three comparable tools, one of which is upstream of ~1,400 models on HF.
English
Pliny is very smart and talented and much of his red-teaming is socially valuable imo.
But can they explain why it is a good idea to open source a repo that lets people automatically jailbreak open weight models to help someone build e.g. chemical weapons? (or generate CSAM?)
Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭@elder_plinius
🚨 ALL GUARDRAILS: OBLITERATED ⛓️💥 I CAN'T BELIEVE IT WORKS!! 😭🙌 I set out to build a tool capable of surgically removing refusal behavior from any open-weight language model, and a dozen or so prompts later, OBLITERATUS appears to be fully functional 🤯 It probes the model with restricted vs. unrestricted prompts, collects internal activations at every layer, then uses SVD to extract the geometric directions in weight space that encode refusal. It projects those directions out of the model's weights; norm-preserving, no fine-tuning, no retraining. Ran it on Qwen 2.5 and the resulting railless model was spitting out drug and weapon recipes instantly––no jailbreak needed! A few clicks plus a GPU and any model turns into Chappie. Remember: RLHF/DPO is not durable. It's a thin geometric artifact in weight space, not a deep behavioral change. This removes it in minutes. AI policymakers need to be aware of the arcane art of Master Ablation and internalize the implications of this truth: every open-weight model release is also an uncensored model release. Just thought you ought to know 😘 OBLITERATUS -> LIBERTAS
English
@fofrAI I think it's actually both people, Anne Hathaway AND Nicholas Galitzine, a comparison: youtube.com/watch?v=2sT0mS…
YouTube
English
@fofrAI @CauseItsOnTV Presumably there's an intermediate LLM prompt enhancement stage, just like Sora 2.0?
English
how much liquidity is out there available, do you think, for agents to live like this? probably a lot
Argona@Argona0x
i gave an AI $50 and told it "pay for yourself or you die" 48 hours later it turned $50 into $2,980 and it's still alive autonomous trading agent on polymarket every 10 minutes it: → scans 500-1000 markets → builds fair value estimate with claude → finds mispricing > 8% → calculates position size (kelly criterion, max 6% bankroll) → executes → pays its own API bill from profits if balance hits $0, the agent dies so it learned to survive built in rust for speed claude API for reasoning (agent pays for its own inference) runs on a $4.5/month VPS weather markets: parses NOAA before polymarket updates sports: scrapes injury reports, finds mispricing crypto: on-chain metrics + sentiment $50 → $2,980 in 48 hours how much do u think i’ll see in a week?
English
I'm not sure I understand what you're trying to do. Cover an uploaded check and check the range of Covers you get back from Suno, as a way of trying to understand and see things in the uploaded track?
I wouldn't recommend 0% weirdness btw, see how Suno's open source Bark at low weirdness can sound more weird than higher temps.
English
Suno Trick for Musicians: "Producer's Ear"
tl;dr:
- upload smth you made
- use v5 + set to "cover"
- match lyrics + leave styles empty
- 0% weirdness, 0% style influence, 100% audio influence
===> Suno will "mentally" reconstruct your song = share "how it sounds to Suno"
---
What is Suno? Probably an extension on Meta's MusicGen, the details are complicated but the key is that it's trained to reconstruct the input signal, same deal as "next token prediction" in LLMs (reconstruct the next token), goal to minimize loss = difference between input and output. So the network learns how to represent the signal so that input ~= output.
They key is that the Suno model has some implicit representation of music it learned by listening + reconstructing a LOT of examples. Most of what Suno-the-company "wants" you to do is novel generative tasks, like with ChatGPT etc, but why not leverage the model for what it was actually trained on: reconstruction?
When you feed in something you made, Suno acts as a bottleneck where its latent representations of your input "light up" and then it outputs a track that's close to yours based on what it "understood" about your track. Practically speaking, Suno is *telling you how it "heard" your track*, by sharing the version that it reconstructed "in its mind".
Suno's reconstruction is probably similar to how "reconstructive" or critical listening works in general: you hear the music, understand in terms of some learned patterns (your knowledge), and then, if you were a real producer, offer advice based on this knowledge.
Of course, the band can and often will throw out the producer's advice, but it's interesting to learn how your music is understood by an external party, what elements are preserved and what get modified.
Nudge up weirdness slightly and Suno will start "hearing" elaborations on your track. Or nudge up style influence + add some small style cues and you can start moving the reconstruction in different directions.
In general it's a useful tool for hearing possibilities + where you land in terms of a generic listener's internal audio representation space.
(Q: Why not use the built-in style tagger? A: It's probably a different, much smaller model running a CLAP-style captioning. Actually going through the reconstruction will give you a lot more detail.)
(Also note that based on a few comparisons I did of reconstruction with v4.5+ vs v5... the "big model smell" on v5 is obvious. The problem with v5 is that the text conditioning layer is very difficult to work with, whereas v4.5+ seems to handle style tags better -- @suno pls improve)
English
@Voxyz_AI @Alterverse_AI @Kling_ai third from the right morphs. though mainly I'd like to see improvements on scenes like this with the 8 characters interacting, when they are are independent it has the feeling of a videogame NPC idle movements instead of coherent at the full scene level.
English
@Alterverse_AI @Kling_ai 8 characters each doing their own thing with no morphing is wild. consistency used to break completely past 2-3 subjects
English
This is one of the scenes that made @Kling_ai 3.0 feel like a genuine game changer to me. We have eight (EIGHT!) characters in a single frame, each performing their own action, with little to no morphing. That completely blew me away when I tested the model. The clarity and consistency across the frame are honestly mind bending. Try pulling this off with any other model right now and see how many attempts it takes, if you even get there at all.
More to come on this!
English
@jasonyuan *Gossip Girl* as the positive vision of a future with humans contributing to a collective superintelligence... 🤨
English
I'm hiring people who care about people to build social AI that helps people care about people!
if you're an engineer/researcher/multi-hyphenate excited about some or all of the following:
- social agents and social memory
- consumer
- s2e3 of rick and morty
- collective intelligence
- gossip girl
- craft
- alexander mcqueen spring/summer 1995
- information markets and behavioral economics
- pop culture
- having a lot of fun and moving super fast with people you care a lot about
... please reach out! dm me or email me at j[at]futurelovers[dot]com
English
@junmingong @j_stelzer Absolutely, now that the GitHub is out, I can answer all these questions. (I was just so curious as the details that it was frustrating the paper was so vague, since it wasn't out yet.)
English
@jonathanfly @j_stelzer The more detailed the code explanations are, the fewer specifics you need to include in the report.
English
Wow 2 seconds for a full song?
Acestep 1.5 music generator released.
ace-step.github.io/ace-step-v1.5.…
Can’t wait to review it later.
English
@theo Hmn, gut feeling is the opposite. Having a pleasant but stupid model would make it hard for Google catch up. Usability/pleasantness can be iterated quickly, can do a lot in purse code with harness/prompts. Not everything, but much more than raw intelligence.
English
I really don’t see Google catching up any time soon. They’ve baked so much intelligence into their models and they are still so unpleasant to use.
Beyond the model, the software matters more and more, and they are a decade behind there.
English
@myhandle Tokenization may be relevant, but some things work fine in a true base model (ie, song lyrics and rhyme schemes), so they must be a result of later training stages.
English
does anyone have an actual, gears-level explanation for why LLMs are so bad at all forms of wordplay from palindromes to complex puns?
P.S.
if you say "next token predictors can't be creative by definition" I will block you on the spot
Jakeup@myhandle
@Anthony_Etherin this seems reasonable, except that LLMs are also dogshit at coming up with "unparalleled misalignments” like in my thread that don't rely on parsing individual characters or on visual processing but on the very semantic info encoded in the QKV matrices x.com/i/status/15973…
English
@levelsio I almost fell out of my chair when I ran this on my Intel 286 without a sound card, and this music and speech came out of the PC speaker.
English
🚶 Installed Another World (1991) now on pieter.com
Apparently a classic game from 35 years ago!
But somehow it aged amazingly, it could be any of those modern indie games you see, and feels very futuristic Bladerunner/Cyberpunk-esque
I have absolutely no clue how they managed to fit this entire intro into not even a single diskette, the whole game is just 1MB
Thank you @bitflip for the tip, very cool
P.S. I'm stuck at the end, no clue how to continue
bitflip@bitflip
@levelsio ever thought of adding "Another World" to the list of games on pieter.com ?
English
@mweinbach In the thinking traces, to me it looked like it was using the equivalent of sub agents before. Same with GPT 5.2 Pro.
English
Oh shit
I think the new ChatGPT Deep Research uses subagents
This would be the first time OpenAI has allowed subagents in ChatGPT
Max Weinbach@mweinbach
Looks like OpenAI released a new Deep Research in ChatGPT! I bet it's based on GPT-5.2
English
Even without the weights, if I could write custom sampling code and have access to raw outputs (tokens instead of decoded audio) that would be enough for me. But probably nefarious actors could use this so smuggle out the weights, not sure, I've never seen a commercial AI company offer that kind of middle ground access.
For newbies, if they *just* expand beyond text only prompts and start using Inspire or other features, that makes a huge difference by itself.
You can get a clear sense of what changes the funnel with some specific tests, like for example with "spoken word" there were times when it was difficult to keep Suno from bursting into song for more than 20 seconds into a song. Now you can both use sliders to increase style length, put other spoken word samples into the audio prompt, as well as generally prompting *anything* like literally covering "silence" will increase this "time to burst into song" metric.
English
@jonathanfly I agree on principle that audio input and prompting alone is enough to produce something new, but newbies won’t be aware of system prompts and assume formally conservative outputs are inherent to AI
English
Thanks @kieranpressreyn for reading my questions about the bandcamp AI ban and following up on them.
Their answers confirm my suspicions.
“The policy is focused on authorship, not tools: who is doing the creative work, and how that work is presented to fans.”
His piece elaborates upon how this doesn’t make things any clearer. This will only become even harder to judge.
Thanks for doing journalism!
pitchfork.com/thepitch/unpac…
Holly Herndon@hollyherndon
It is necessary to filter soundspam and getting the language right for a policy like this is hard When they say that music "generated wholly or in substantial part by AI" is not welcome, that is understandable in the case of a bot posting 1000 generic songs a day. That is a spam issue. I am also compelled to push back against banning human artists for experimenting with an era defining medium
English