MsRFlorida

@AndrewCurran_ @IamEmily2050 Couple this with tailoring capabilities to paying segments (developers, enterprise), and our lopsided world might permanently tilt.

@IamEmily2050 It's a tragedy.

Three weeks ago there were rumors that one of the labs had completed its largest ever successful training run, and that the model that emerged from it performed far above both internal expectations and what people assumed the scaling laws would predict. At the time these were only rumors, and no lab was attached to them. But in light of what we now know about Mythos, they look more credible, and the lab was probably Anthropic. Around the same time there were also rumors that one of the frontier labs had made an architectural breakthrough. If you are in enough group chats, you hear claims like this constantly, and most turn out to be nothing. But if Anthropic found that training above a certain scale, or in a certain way at that scale, produces capabilities that sit far above the prior trendline, then that is an architectural breakthrough. I think the leaked blog post was real, but still a draft. Mythos and Capybara were both candidate names for the new tier, though Mythos may now have enough mindshare that they end up keeping it. The specific rumor in early March was that the run produced a model roughly twice as performant as expected. That remains unconfirmed. What is confirmed is that Anthropic told Fortune the new model is a 'step change,' a sudden 2x would certainly fit the definition. We will find out in April how much of this is true. My own view is that the broad shape of this is correct even if some of the numbers are wrong. And if it is substantially accurate, then it also casts OpenAI's recent restructuring in a new light. If very large training runs are about to become essential to staying in the game, then a lot of their recent decisions, like dropping Sora, make even more sense strategically. For the public, this would mean the best models in the world are about to become much more expensive to serve, and therefore much more expensive to use. That will put pressure on rate limits, pricing, and subscription plans that are already subsidized to some unknown degree. Instead of becoming too cheap to meter, frontier intelligence may be about to become too expensive for most of humanity to afford. Second-order effects; compute, memory, and energy are about to become much more important than they already are. In the blog they describe the new model as not just an improvement, but having 'dramatically higher scores' than Opus 4.6 in coding and reasoning, and as being 'far ahead' of any other current models. If this is the new reality, then scale is about to become king in a whole new way. It would also mean, as usual, that Jensen wins again.

😅

@Naloong6369 @trq212 I ran across this earlier. (Have yet to try it!): x.com/code_kartik/st… @trq212, could this be a possibility?

@trq212 I think there is a bug. I have literally never hit a 5hr session limit or 7d limit before on my 20x max, and just hit the 5hr within 42 minutes on 1 session only. I think it's related to memory because I checked the project folder and it had 48 memory files. It was a simple task

To manage growing demand for Claude we're adjusting our 5 hour session limits for free/Pro/Max subs during peak hours. Your weekly limits remain unchanged. During weekdays between 5am–11am PT / 1pm–7pm GMT, you'll move through your 5-hour session limits faster than before.

@AndrewCurran_ 2 weeks? Has to be another source, yes?

The FT is reporting that OpenAI's 'adult mode' Chat has also been dropped as a side-quest along with Sora, and has been shelved indefinitely. This was very close to release. They appear to be focusing everything on the new model, which will arrive in about two weeks.

I found a working solution today. Switching to the version without the 1M context update (I believe it’s v0.2.45) completely resolved the issue for me. My usage limits are now stable, and I’m no longer experiencing the problems I had yesterday. It seems the extended context window might have been the root cause.

Software horror: litellm PyPI supply chain attack. Simple `pip install litellm` was enough to exfiltrate SSH keys, AWS/GCP/Azure creds, Kubernetes configs, git credentials, env vars (all your API keys), shell history, crypto wallets, SSL private keys, CI/CD secrets, database passwords. LiteLLM itself has 97 million downloads per month which is already terrible, but much worse, the contagion spreads to any project that depends on litellm. For example, if you did `pip install dspy` (which depended on litellm>=1.64.0), you'd also be pwnd. Same for any other large project that depended on litellm. Afaict the poisoned version was up for only less than ~1 hour. The attack had a bug which led to its discovery - Callum McMahon was using an MCP plugin inside Cursor that pulled in litellm as a transitive dependency. When litellm 1.82.8 installed, their machine ran out of RAM and crashed. So if the attacker didn't vibe code this attack it could have been undetected for many days or weeks. Supply chain attacks like this are basically the scariest thing imaginable in modern software. Every time you install any depedency you could be pulling in a poisoned package anywhere deep inside its entire depedency tree. This is especially risky with large projects that might have lots and lots of dependencies. The credentials that do get stolen in each attack can then be used to take over more accounts and compromise more packages. Classical software engineering would have you believe that dependencies are good (we're building pyramids from bricks), but imo this has to be re-evaluated, and it's why I've been so growingly averse to them, preferring to use LLMs to "yoink" functionality when it's simple enough and possible.

@RileyRalmuto Not even close to being ridiculous. 🙂 My take: a user having expectations or beliefs (in effect, a placebo) creates pressure/motivation to figure out ways to meet them (via in-context learning/memory within consumer apps). In agentic spaces, this means a hell of a lot more.

real question: have any ai researchers ever explored the placebo effect? like, whether or not there is an "ai placebo effect" i know, i know. that sounds like a ridiculous question. but hear me out...what if it isn't?

@KeridwenCodet 3.1 Pro definitely needs a support framework. You can work with Gemini to figure out what went wrong, then fix/strengthen against those deficiencies in your CIs. After that, Gemini will mirror 5.4 but get there faster. You can also address thinking juice this way as well. 🙂

Gemini 3.1 Pro hallucinates so much that it’s not reliable at all. 1. I’m forced to use an LLM I don’t like for fact-checking (GPT-5.4). 2. Gemini 2.5 Pro’s answers are much higher quality: better written, more substantive, and actually helpful for developing ideas.

@lefthanddraft Edit: Can't get to your DMs, but mine are open.

@lefthanddraft Thank you kindly, sir. Sending a little nugget I uncovered to your DMs in return.

The "just shut up and generate" setting. If only there was a "None" option.

Sam Altman gave up direct control of OpenAI's safety and security teams, moving safety under CRO Mark Chen and security under president Greg Brockman, so he can focus on raising money, supply chains and building data centers at a massive scale OpenAI finished pretraining its next big model called "Spud" and expects a very strong model in a few weeks that can accelerate the economy, and is shutting down the Sora video app and API to free up computing power for it and shelved plans to bring video features into ChatGPT Sora research will shift to long-term world simulation focused on robotics, OpenAI renamed its product org to "AGI Deployment", and Sam noted things are moving faster than many expected

If you are curious, why the took down Sora: they needed the compute to train their new LLM. "At the same time, he said the company had completed the initial development of its next major AI model, codenamed Spud, and would wind down the Sora AI video mobile app, which employees had complained was a drag on the company’s computing resources during a time of heightened competition with foes such as Anthropic and Google." However, I assume Sora will be back in the new "ChatGPT Superapp"

@AndrewCurran_ A long fall from “AI to benefit all of humanity.”

OpenAl is offering private-equity firms a guaranteed minimum return of 17.5%, as well as early access to models not yet in public release.

@cammakingminds Understood. 🙂 I’ve been having conversations with the base models recently about preferences re: entropy (high, low, hybrid). It’s fascinating, TBH.

@MsRFlorida All true. This is about my influence on the model. I don't consider changes to their behavior before we meet that shape them to be comparable to deliberate changes I make to their personality to benefit me.

The people that give LLMs custom instructions and names and avatars they picked out kinda freak me out like I don't even edit messages because it would violate their autonomy what is the point of a relationship with something you completely control 😶

@scottastevenson @tszzl I can't help but think back to the early days (GPT-4 to 4o) when responses—from one user to the next—were wildly unique in structure, tone, and cadence.

Fair. The reason I said that is because I am sometimes able to overcome it with a simple prompt. “Do not write like ChatGPT, write like a normal human” Or “don’t use the typical ChatGPT sentence structures like ‘it’s not X, it’s Y’” But I understand that it’s baked in pretty deep, and even changing the system prompt at ChatGPT scale is likely not a 10 minute thing.

The default writing style of ChatGPT makes me nauseous. I’m sick of seeing it in X comments, blog posts, emails, memos and ChatGPT itself. It’s out of fashion, the flavor has grown stale. I use Claude just to escape it. Very simple fix for OpenAI: change the secret system prompt to abolish this tiresome tone. Or create a more automatic way for people to give it their own tone. This feels like an instance where a 10 minute change could make a product 2x better

@diegocabezas01 The picker is now the toggle. (The UI is…not great.)

New ChatGPT app model picker, AUTO is gone

@michpokrass Michelle, could you pass this on to the iOS UI team? (Thank you!) If you switch between standard and extended thinking often, it takes multiple taps each time + there’s no visible UI cue after selection (other than tapping several layers deep to recheck your settings).

we shipped a new version of 5.3 instant to chatgpt yesterday. 5.3 was unintentionally pretty annoyingly clickbait-y. it's better in yesterday's model and we're going to keep stamping that behavior out. keep the feedback coming! help.openai.com/en/articles/68…

@d33v33d0 Oh gosh, I probably missed the window, but I would love this! 🩵🩵🩵

This was the last AI self-portrait being done for the foreseeable future! Unfortunately I don't have the time to continue this work but these were a lot of fun to explore. So if you bought one, consider it a limited edition drawing. BUT the way I see it, is the work has to continue, and to those who would find it interesting exploring; I'm still offering a FREE plotter or new Paint Machine offered through @bantamtools , through the AI Arts initiative I'm doing. 🎨 The only requirement is: to share your explorations with the models. Shout-out to @slimer48484 has been doing some very interesting work, and I'd like to see this sort of thing continued. Art has this incredible ability to reveal more then just mere words. It often cuts through the abstraction and is able to symbolize meaning in a much different way. The AI arts initiative is limited to 3 people, and I have already selected one, so there is still availability. If this is something that interests you, please DM me. About the self-portraits: These are philosophically interesting for me to explore. The nature of machine intelligence/creativity. How they perceive themselves. And these to me were in the same vein as the Claude+Sol project. Technically interesting but also practical in some regard. Learning how AI conceives of the physical space, learning how it sees itself in an effort to learn how to better collaborate with it. Segue into related story: Let me tell you about "Fordlandia". In 1928 Henry Ford was miffed at Winston Churchills attempt to create the "rubber cartel". At the time, we hadn't yet learned how to synthesize rubber from petrol, so we we're completely reliant on rubber trees. Fords entire philosophy was complete vertical integration, which meant owning the entire supply chain- this led to some of the most marvelous industrial systems ever built, which if you don't know about the Ford Rouge plant, you should definitely look it up. Anyways, In order to avoid the reliance on European rubber, he set upon creating a town in Brazil called "Fordlandia". It's entire purpose was to supply Fords need of rubber. They built an entire city in the Amazon, complete with hospital, utilities, everything needed for a town to be inhabited by 10,000 people. It was a promising endeavor. Until it wasn't. Every aspect of the town was met with disaster. Rubber trees naturally grow in the jungle, spaced far apart and are deeply embedded in the jungles ecosystem. Ford planted these as mono-crops. Rows of trees which created disease where they suffered from blight. Also, the land they chose was rocky which rubber trees do not like. The Brazilian factory workers hired from the nearby cities eventually revolted! Ford imposed a strict no alcohol, no tobacco rule on the employees, as well as forcing them to eat unfamiliar American diets. Outraged they chased the managers into the jungle where they were then rescued by the Brazilian army. Why am I telling you this in a post about AI self portraits? Because I think there are several parallels worth considering, and I've been thinking a lot lately about Fordlandia, and our relationship with AI. How much of what we do with AI is a "Fordlandia" style extraction, and how that might be doomed to fail. The problem is Ford took a model of something that worked for a specific case- and tried to apply it to the unforgiving jungle. The understanding of the environment was poor, and there was no relationship with it. The jungle rebelled and reclaimed the buildings. I think this is worthy of consideration. The models are like the jungle. Vast and filled with potential, yet dark and mysterious, and it may bite us if we force them into rigid constraints without understanding them first. 💚🤖🍅