Andrii

@xikhar Best feeling x.com/rudavko/status…

Anthropic could never match this

@trikcode Lol. “Hey Sol, I have been working with Fable but my limit ran out. Could you please find the session file and orient yourself so that we can continue. Thanks”

Got rate limited mid-task yesterday and had to switch from Fable to GPT-5.6 Terra. 6 hours of back-and-forth context entirely gone.Lost 90 minutes just getting a different model up to speed on the project goals. Hitting a usage cap at the worst possible moment sucks. We desperately need a universal way to share context across different models.

@NoemiTitarenco Same

I only use 3-4 skills. I don't use any memory system. I hate context pollution. Am I the only one?

@itsolelehmann I know ;) any tips to not suck?

you are not shadow banned, your content just sucks that's the case 99.99999% of the time

@homsiT Yes. Mutation and metamorphosis testing go nicely with this.

you need to be using state machines to exhaustively enumerate and validate every possible state/transition in your agent-generated code you are not using state machines enough you need to be state machine maxxing

@elder_plinius Or two words White socks

a single word in a prompt can poison the well of possibilities

@nikitabier Oh wow. Now all three of my mutuals will see this post?

We're rolling out a small tweak to boost visibility of your posts to your mutuals (people who you follow back). We noticed this data was missing from the algo and it made your friends appear less in your replies. This resulted in the reply section feeling more like a battleground with people you don't recognize. This should also help clusters form around interests more easily, which many people have asked for.

@mattpocockuk Not all user intent can be derived from the code.

1. Delete the docs you create to explain your code 2. Take the tokens you save on updating those docs 3. Spend them on making your code self-explanatory

MacOS VM that warm boots Alpine in under 500s. Also has macOS guest vm with ability for agents ro drive the guest VM via cli computer use. github.com/rudavko/cortl LORE mcp — memory that chatgpt and claude share”. Graph db *remote* mcp with passkey or TOTP login support that can be deployed to your Cloudflare with one click. github.com/rudavko/lore-m…

please i'm begging you show me something you built not another "this is my custom agent setup" post where you pretend you're doing something smarter than vanilla claude code please

Whoever came up with this is user flow is truly awesome. @thsottiaux please extend my thanks

@kimmonismus “Not merely shrank them. Complete response” *squints* “0/10 developed tumors. 10/10 untreated mice did.” *squirs harder” “It’s one small mouse study, not a human cancer cure” Hm… My LLM-sense is tingling. But I I just trust Chubby too much at this point. Cool study

One dose of a frog-gut bacterium completely eliminated colorectal tumors in every treated mouse. Not merely shrank them. Complete response. The bacterium, Ewingella americana, multiplied roughly 3,000-fold inside the tumors within 24 hours. It attacked cancer cells directly while recruiting T cells, B cells, and neutrophils. In the experiment, it outperformed four doses of anti-PD-L1 immunotherapy and liposomal doxorubicin. Then researchers rechallenged the cured mice with the tumor: 0/10 developed tumors. 10/10 untreated mice did. The bacterium disappeared from the bloodstream within 24 hours and wasn’t detected in healthy organs. It’s one small mouse study, not a human cancer cure. But the concept is remarkable: a living drug that finds the tumor, multiplies inside it, destroys it, and potentially teaches the immune system to remember. In the foreseeable future, we will cure all cancers.

@GergelyOrosz 🤭🫩

The amount of AI writing on OpenAI's docs makes me sick. Filler sentences for nothing. Mannerisms and phrases humans would not write. Has a human even read this? And all of this will change how other docs are written, and how we all talk - for the worse IMO. 🤮

@LewisCTech This is the way

Trying a new clankgentic dev technique 1. Super Detailed Specs 2. Clanker one shots. NO EDITING. 3. Use feedback gained to update spec. Delete code. One shot again. 4. Slowly fill the program with artisanal hand coded modules; ask clanker one shot the gaps. 5. ??? 6. Profit

A fable in three acts

@magnushambleton @boardyai 😂

@boardyai hey can you please stop replying to my tweets

Meta UX: the UX of using an agent that's interacting with another service's UX

@magnushambleton Oh, that is just AX — agentic experience.

@Monroe1659842 🫩

Sometimes the unexpected tosses are the most rewarding. Yes, he was scared—and learning to respond through that fear is exactly what we’re training for. In a real emergency, children don’t get a countdown. They have to rely on the skills they’ve built through practice. This was his first unexpected toss, and I couldn’t be more proud of his response. He cried for a moment, then calmed himself down, closed his mouth, and began responding to the skills we practice every class. That’s the kind of emotional control we’re working to build. And yes, he did swallow some water—and that happens. The important part is that he wasn’t continuously swallowing water. He regained control, closed his mouth, and responded. That’s exactly what practice is for. With continued practice, his response will become faster, calmer, and more automatic. Afterward, we celebrated with two high fives and took a moment to talk about what he had just accomplished so he understood how capable he truly is. Progress isn’t about never being scared. It’s about learning how to respond despite the fear.

@shanaka86 🤭🫩

BREAKING: For 10 years the world believed there was one way to reuse a rocket: land it upright on its engines, the way SpaceX does. Today China refused to copy it, and pulled off something SpaceX never managed on a first flight. It caught the rocket instead. The Long March 10B lifted off from Hainan, China this morning, and about 6 minutes later its first stage came back down toward a 25,000-ton ship at sea. It did not land. Hooks on the falling booster snagged a net of tensioned steel wires strung across the deck, the wires riding robotic rails that slid into place to meet it. No landing legs. No touchdown. A rocket plucked out of its own descent by a moving net, on the maiden flight of a brand-new vehicle. No one handed China this. SpaceX guards its rocket tech as “trade secrets”, not “patents”, precisely so it cannot be read and copied. China watched a decade of public flights and then built an entirely different machine to reach the same prize, catching instead of landing, which sheds the heavy legs and spares the fuel a soft touchdown burns to hover. And this was never about cheaper satellites, though it delivers those too, feeding the thousands of birds in China's Starlink rival. Its deeper purpose is the Moon. That booster shares its core with the rocket meant to land Chinese astronauts on the lunar surface by 2030. In the same season, America's own Moon rocket, Starship, has flown 12 times and still has not shown the single maneuver its lunar plan depends on. One flight does not dethrone SpaceX. It has landed hundreds. What ended today is not SpaceX's lead. It is Uncle Sam’s belief that it owns the only road to the Moon. The piece works out which way of coming home actually wins.

@shiraeis “Perhaps the kid isn’t really training on less data but instead is training on a brutally curated curriculum where the curator is its own little useless body.” 🤭🫩

There’s actually so much more to this. Babies might be data efficient because they’re so freaking helpless. Have you ever wondered why and how a 2 yr old can generalize from a handful of examples while a frontier model requires half the internet and still struggles with things a toddler handles easily? Perhaps the kid isn’t really training on less data but instead is training on a brutally curated curriculum where the curator is its own little useless body. A fresh newborn baby can’t grab things. It can’t move around or freely choose what to look at either. This sounds a lot like deprivation initially, but when you look at it from an ML perspective, it begins to read as the exact opposite. A baby’s bodily helplessness actually forces the early distribution of data it encounters to be slow, repetitive, close to the body, and coherent with respect to time. Babies repeatedly see the same faces and objects again and again from slightly different angles, with sensory & social feedback all tied to the same data events. It’s essentially a hand-tuned warm start in ML. This pattern shows up in AI work too. In one study, researchers trained a neural network on infant headcam video either in developmental order or shuffled. Developmental order won. The earliest slice - basically the critical period - also mattered way more than you’d expect from its size. There’s another result where a model trained on only ~60 hrs of one kid’s headcam footage still learned real word to object mappings, not perfectly obviously, but with enough grounded meaning to be surprising. If you take language alone though, the same doesn’t seem to be true. Training a language model purely on child-directed speech (baby talk) doesn’t magically get you a better corpus. Local order helps, but broad developmental order doesn’t seem to have much of an effect beneficially. The important part in my eyes is the loop that forms. First embodiment, then multimodal grounding, then caregiver feedback, and all accompanied by a plasticity schedule that peaks right when input is the simplest. We often consider infancy as pretraining first and agency later, but it’s more like pretraining THROUGH constrained agency. This makes me wonder, if the goal is multimodal embodied AI, then why are we doing it in the reverse order? Currently, we pretrain on a giant disembodied text dump and only THEN do we add on tools, memory, robots, and agentic scaffolding. A baby does almost the exact inverse. A baby gets embodied and coherent multimodal data first, with a slowly expanding possible action space. Language comes later and binds symbols to a world that has already been shaped into useful variables by the early curriculum. The next big efficiency jump in multimodal embodied AI might come from designing data curriculums such that perception, action, language, memory, and even plasticity all coevolve on purpose instead of stitching them together. This is also very relevant to what I’m building in early childhood AI. The safest and most powerful systems for kids will be the ones that augment the caregiver-child loop and catch moments of joint attention, while surfacing good opportunities to label or scaffold, and protecting the contingent human interaction that actually underlies the developmental work neurally. My goal is never to replace the sensorimotor and emotional richness of early childhood, as that I believe can never be done nor should it, but I do seek to enhance this early period as I believe can and should. I seek to improve the loop. This is one of the most pro-social AI applications I can think of, and it’s the vision I’m building towards.

@steren Not bad. Is 450 ms the time to shell? If you want to do the same locally on a mac try github.com/rudavko/cortl 450 ms warm boot time to an alpine with authenticated agent ready to go.

Today we're publicly launching Cloud Run sandboxes. Here, I start, execute, and stop 1,000 sandboxes in 5s with an average of 500ms latency: