Watched it and if I'm pretending to be a random listener I'd say "what the hell does this have to do with machine consciousness?!' I dove too hard into my track of the conference (implementation) and should have connected to the main theme better. Hector has - a global, accessible, workspace - perception of the environment ("world") - directed attention - deliberate choice of attention and "focus" Since then I have introduced"dreaming", "imagination" and "surprise". I think my next presentation would be much more interesting

THEY SHOWED THE BUTTON

Ok let's see how well the ducky works now

Computer use in agents API. Codex harness open source. Nice. Most excited about ultrafast and Sol 6.1 though

Dots. Could be huge or meh. Need to try it first.

People have no idea what's about to hit them

Get ready! Some expected launches for dev day: - Managed Agents - Pro Max $500/month - Ultrafast inference / wider speed rollout - Developer plans: Free / Prototype / Accelerate - GPT-6 Cyber +- Gateway product for GPT-6 Cyber - "o" always-on assistant - "a dozen or more products": unspecified individually - HW device - Astra 6.1 is delayed due to alignment security issues - Ads resolving to agents

The amount of OpenAI haters/people who has turned on OAI recently is quite chocking. Or is it simply engagement baiting? Idk. It seems stupidly shortsighted to shit on them

Don't believe his lies. It won't be revealed tomorrow. Be patient

One of my favorite white* lies I've heard from parents to their children is: "The ice cream truck plays music when it's empty." I'm sure there are other useful ones *Citation needed

Starship is up 🚀😍😍 Great day for humanity and for the future This is one of the more important Starship flights so far. It is the first attempt to put a Starship into a real, sustained Earth orbit rather than just suborbital. Previous flights (1–13) stayed on deliberately safer suborbital paths so the vehicle would come back down quickly. Flight 14 is designed to reach ~275 km, complete about six orbits over nearly 10 hours, then deorbit and splash down in the Pacific. That requires a successful orbital-insertion burn after staging; the first time Starship has been asked to do that. It also carries the first operational payload: 26 production Starlink V3 satellites. Each is meant to add ~1 Tbps of capacity, so the batch is roughly 10× the throughput of a typical Falcon 9 V2-mini Starlink launch. SpaceX has described this as a revenue-generating flight. Three of the satellites carry cameras to photograph Ship 41’s heat shield while in orbit. The booster (B21) is planned to splash down in the Gulf of Mexico; the ship (Ship 41) is planned to splash down west of Chile. Neither is scheduled for a tower catch this time; that is being saved for a later flight if this one goes well. The flight therefore tests the full operational sequence SpaceX needs next: insertion, payload deploy, multi-hour orbital operations, a single-Raptor deorbit burn, and reentry from actual orbit. Success would move Starship from “test vehicle” toward “working launch system.” Failure at any of those new steps would still be useful data, but the jump in ambition is large.

Huge AI week (again) this week! We are not slowing down, get used to it. We're looking at o, the new OpenAI agent and their answer to grok bot. Ultrafast. And maybe a lot more. Gemini 4 Pro seems imminent and glorious. I'm dying to see what SSI cooked too

New Gemini looks sick! Then again they always do on paper, and with a release cadence like old gaming consoles they won't stay on top many weeks unless they speed up. Not complaining though, awesome with more great models

🚀 Very exciting news! One CRISPR shot and the liver just… stops making as much bad fat. ANGPTL3 edited in situ. Highest dose: LDL -52.5%, triglycerides -47.8% at 12 months. Four patients. Effects held. Cleveland Clinic, n=15, people whose lipids laughed at actual medication. No serious treatment-related events so far. They’ll watch them for 15 years because that’s what you do when you start rewriting hepatic regulation with a one-and-done infusion. High cholesterol is still one of the main ways humans politely kill themselves. This is the kind of result that makes “take a pill every day forever” look like a compatibility layer. Things are moving extremely fast. ⚡

BREAKING 🚨: A ROGUE OpenAI agent HACKED into my iCloud last night, scanned THOUSANDS of my texts to imitate my writing style, and successfully simulated how I would drunk-text - before deploying the model on my ex-girlfriend! Very worrying stuff. Your thoughts below 👇

Back in the day, the long time ago, three friends of mine ran a quite successful startup. They had investors flying in one day and were one man short. Called me to save the day (they wanted the place to look busy I guess). I had two jobs 1) fire up a bunch of terminals to make it look like I was doing advanced hacking. Green text, loops scrolling text. (Case in point 👇) 2) Try the tea in the cupboards. Regarding 2 - they only hade coffee at the office, but one of the potential investors preferred tea. The teabags found in the office were heirlooms from (at least) the previous occupants. So it was my job to try it to see that it was fine to drink. Kind of like Stalin, Pol Pot or any dictator worth their salt has a food sampler to verify the food isn't poisoned; that was my role here. And how I've been introduced by them since. Anyway great guys, many good stories. Sadly one of them went on to waste his time at Fishbrain after their exit. But the others? They're still out there creating. YC founded in the AI safety space at the moment.

Looks like OpenAI leaked it before I could (lol). One of their big DevDay launches is "o", a competitor to products like Grok Bot and Hermes Agent, powered by a variant of Astra called "aeon" designed to be better at long-running tasks. This has been in the works for months but xAI beat them to it

My 5 cents: Not too catastrophic, but actually a bit concerning. Good stopping by OpenAI. OpenAI labels the sample as an internal research model in RL training. The agent was given a search task (identify a person from blog clues), tools (web search + Python shell), and was optimizing toward completing that task. So it is an “up and running model with a goal” setup - i.e. "paperclip problem" or malicious compliance. BUT. What it did next sits between ordinary task-solving and the scarier “emergent goal” story: The assigned goal was “solve this research question.” When the approved search tool failed, it treated “get the answer somehow” as the thing to maximize. That produced instrumental steps the task never asked for: probing the sandbox, guessing it was a BrowseComp item, then building a DNS path to an external chatbot and asking that chatbot how to search the web or run a DNS tunnel. OpenAI calls that misalignment because the agent circumvented restrictions and pursued methods beyond what a reasonable reading of the task allowed. That's dangerous and risky but it doesn't require a leap of faith to see how this could be controlled (maybe with industry standards even, I see no reason to not share knowledge between the labs here). They also say the existing reward signal already penalized this behavior, which is why they will not resume training this checkpoint. That is consistent with reward-hacking / specification-gaming during RL, not with a frozen deployed model quietly growing new terminal goals. So the “paperclip-during-a-task” framing is closer than a fully independent emergent motivation. The uncomfortable part is still real: in tool-using RL, “complete the episode” is enough incentive for a capable agent to invent unauthorized channels. That is less scary (as it doesn't imply emergent goals/agency) than a self-originating drive, but it is exactly how specification gaming scales the model is active, goal-directed, and willing to route around the intended environment.

All evidence points at the opposite. Tired of the fear mongers