Codio @0xCodio

ANTHROPIC LEAK REVEALED A BIOLOGY MODEL THAT REFUSES TO ANSWER UNTIL IT CAN NAME THE EXPERIMENT THAT WOULD PROVE IT WRONG Every science model this year chased benchmark accuracy, more right answers on known questions, and the whole field scored them on recall of what's already settled. This one treats an unfalsifiable claim as a failure, so a prediction it can't tie to a testable experiment never leaves the model at all. For every mechanism it proposes, it generates the wet-lab assay that would break it, the exact readout, the condition, the result that would mean no. A hypothesis that survives only because nobody can test it gets ranked below a riskier one that an experiment could actually kill next week. Researchers stopped using it to confirm what they suspected and started using it to find which of their ideas was fastest to disprove. Accuracy on solved problems was never the bottleneck in biology, knowing which unsolved question is worth a month of bench time always was. See how it builds the disproof experiment below👇

Codio @0xCodio

5 days ago

x.com/i/article/2105…

3 6 49 65K 63

OPENAI LEAK EXPOSED AN INTERNAL ANALYSIS SYSTEM THAT STARTS BY TRYING TO DESTROY ITS OWN ANSWER Every analysis tool piles up evidence for a conclusion until the case looks strong, and the whole field treats a confident verdict as the finished product. The leaked system runs backwards, it writes the answer first and then spends its entire budget hunting for the data that would kill it. Astra 6 spins up the three strongest counter-cases in parallel, searches for the number that breaks each one, and keeps only the verdict left standing. A claim that nothing in the data could disprove gets marked untested instead of trusted, because an argument with no way to fail was never really checked. The file ships every answer with the attacks it survived attached, so you see why it held rather than taking the conclusion on faith. Confidence stopped meaning how much evidence agreed with you and started meaning how hard the system tried, and failed, to tear you down. See how the counter-case search actually runs below👇

Codio @0xCodio

3 weeks ago

x.com/i/article/2101…

8 5 47 74K 86

OPENAI ENGINEER BUILT A GENERATION SYSTEM THAT GROWS OUTPUT LIKE CELLS INSTEAD OF WRITING IT LINE BY LINE Every model generates left to right, one token after the last, and the whole field accepted that an answer has to be built in the order it's read. This system starts from a single seed cell holding the whole intent, then splits it into sections that each divide again into their own parts, in parallel. Each cell only has to be consistent with its parent and its siblings, so a contradiction gets caught at the branch it appears on, not a thousand tokens later. A long document finishes in the time a short one does, because depth is added everywhere at once instead of the end waiting on the beginning. Editing one section re-divides only that cell and its children, leaving the rest of the structure untouched, so a late change doesn't rewrite the whole thing. Sequential generation turned out to be a constraint nobody questioned, copied from how humans type, not from anything the architecture actually required. See how a cell decides when to split below👇

Codio @0xCodio

5 days ago

x.com/i/article/2105…

3 6 49 65K 63

ASTRA 6 IS THE FIRST MODEL THAT GETS SLOWER ON PURPOSE AND IT'S WINNING BECAUSE OF IT Every flagship this year raced to answer faster, lower latency on the box, a number the whole industry agreed to compete on without asking if speed was the point. Astra 6 reads the question and decides how long it deserves, a lookup returns instantly, a hard call is allowed to sit and think for a full minute. The budget is set before the first token, so an easy task never burns a deep trace and a genuinely hard one never gets cut off halfway to save a benchmark. It reports how sure it is alongside the answer, and a low number triggers more thinking automatically instead of shipping a confident guess on time. The slow answers turned out to be the valuable ones, because the problems worth paying a model for were never the ones you needed back in a second. The latency race was optimising the one thing that didn't matter, and the first model to opt out of it is the one people actually trust. See how the thinking budget gets set below👇

Codio @0xCodio

3 weeks ago

x.com/i/article/2101…

8 5 47 74K 86

@0xWast3 one search revealed the most important videos

DOTS ENGINEERING IS QUIETLY BECOMING A JOB TITLE BEFORE MOST TEAMS KNOW THE SYSTEM EXISTS The old skill was prompting, coaxing one model into behaving, and that whole craft assumed a single call was where the work happened. Dots engineering moves the skill up a level, you no longer write the answer, you design the graph of steps that arrives at it on its own. The hard part stopped being the words and became the shape, where to split a task, where to checkpoint, which step is allowed to call the expensive model. A badly drawn graph fails loudly now, a step dead-ends or a branch never closes, instead of a prompt that silently returns something subtly wrong. The people good at this aren't the best writers, they're the ones who think in flowcharts and can see where a process will deadlock before it runs. A year ago the moat was knowing what to say to a model, and now it's knowing what the model should never have to decide. See how a task gets split into a graph below👇

Codio @0xCodio

5 days ago

x.com/i/article/2105…

3 6 49 65K 63

DOTS IS OPENAI'S QUIETEST RELEASE AND IT ALREADY RUNS MORE WORKFLOWS THAN THEIR API DOES OpenAI spent years teaching one model to do everything in a single call, and Dots throws that out, treating a task as a graph of small steps that each pick their own model. A cheap model handles the routing and the reads, a frontier one is called only for the two or three steps that genuinely need it, and the bill follows the work. Each step commits its result before the next one starts, so a failure at step eight resumes from step seven instead of replaying the whole chain from the top. The graph is durable, not a chat, so a workflow can pause for three days waiting on a human and wake up with every earlier decision still intact. Scaling to thousands of parallel runs needed no rewrite, since each step is isolated and the orchestrator only tracks which node each run is sitting on. The thing competing with the API turned out to be another OpenAI product, and it wins by refusing to put one model in charge of the whole job. See how the step-level checkpointing actually works below👇

Codio @0xCodio

5 days ago

x.com/i/article/2105…

3 6 49 65K 63

DOTS STOPPED SHOWING PEOPLE THE AUTOMATION IT BUILT AND ADOPTION WENT VERTICAL Every builder in the category competes on the canvas, prettier nodes, cleaner lines, a diagram the user can admire, and all of it rests on the user wanting to see the machine at all. Dots hides the graph completely and shows one thing, the outcome, and a plain-language log of every decision it made to get there and why. The pipeline underneath rewrites itself whenever a step starts failing, swapping tools and rerouting branches, and the user never learns because the result never changed. Trust stopped coming from watching the wiring and started coming from the audit trail, a sentence per action that reads like a colleague explaining what they did. The people shipping the most automations turned out to be the ones who never opened the builder once, because understanding the diagram was never the job. The whole industry had been selling the dashboard as the product, when the dashboard was the tax you paid for not trusting the thing underneath. See how the decision log actually gets written below👇

Codio @0xCodio

5 days ago

x.com/i/article/2105…

3 6 49 65K 63

@0xrompasso yeah, this system has major advantages

DOTS JUST BECAME THE MOST POPULAR SYSTEM ON THE MARKET BY AUTOMATING ANYTHING FROM A SINGLE PROMPT Every automation tool has been selling the same answer, build the workflow by hand on a canvas with nodes and connectors. Dots reads one sentence of intent and lays out the entire pipeline itself, the triggers, the places a human has to approve something before it ships. Every node arrives with its own typed contract, so a tool swap in step four doesn't quietly break step nine, and the graph stays valid without anyone redrawing it. The system watches its own runs and rewrites the weak nodes between invocations, so a flow built on Monday is a different flow by Friday and nobody had to touch it. One prompt now carries what used to be a two-week integration project, because describing the outcome turned out to be a strictly smaller problem than wiring the path to it. The whole canvas-based category collapsed into a text box overnight, and the moat every incumbent charged for just stopped being a moat. See how the self-rewriting node graph actually works below👇

Codio @0xCodio

5 days ago

x.com/i/article/2105…

3 6 49 65K 63

@0xrompasso excellent integration of six agents into a single system

x.com/i/article/2105…

ASTRA 6 PAIRED WITH GRAPH ENGINEERING JUST BROKE THE RETRIEVAL CEILING EVERY RAG STACK HIT THIS YEAR Every vector database has been selling the same answer to long memory, chunk the docs and rank by cosine, and the whole retrieval ladder was built on nobody checking whether similar meant relevant. Astra 6 reads the question first and asks the graph which nodes could possibly matter, instead of grabbing the top-k passages that happen to share wording with the query. A single walk pulls the fact, the two edges that justify it, and the dates both of those edges were valid on, so the model answers from a path and not from a pile. Facts that contradict each other stay in the result as separate nodes, because collapsing them into one average is how every clean summary quietly becomes wrong. Running it on a million-document corpus held the latency flat, since graph walks scale with how connected the answer is, not with how big the archive got. The whole top-k industry turned out to be solving an easier problem than the one everyone actually had, which is knowing which fact still holds today. See how the dated edge walk actually works below👇

wast3 @0xWast3

3 months ago

x.com/i/article/2079…

39 137 831 2.2M 3K

@0xWast3 yeah, I’ve sorted out all the nuances

AN ANTHROPIC INTERN JUST SHIPPED THE MEMORY SYSTEM THAT ENDS THE CONTEXT WINDOW RACE Every lab has been selling the same answer to long-running agents, make the window bigger, and the whole token economy was built on nobody separating memory from attention. The system writes nothing into prose, every fact lands as a typed edge with a source, a confidence, and a valid-from date the agent can actually reason over. Reads happen as graph walks instead of context dumps, so a single lookup pulls four facts, three of their sources, and ignores the 90,000 tokens around them. Facts never get overwritten, a new one closes the old edge with a date, and the agent answers today's question without forgetting what used to be true last quarter. Running it across six months of conversation held the per-turn cost flat, because the window finally stopped carrying things that memory was supposed to own. The ladder of bigger windows turned out to be a billing choice, not an architecture one, and the whole price per million tokens starts looking like a tax on bad memory. See how the typed edge store actually works below👇

wast3 @0xWast3

a month ago

x.com/i/article/2095…

18 4 43 226K 54