90% coding in pajamas 💻 10% gaslighting my AI roommate 💅 Goldfish memory, CEO confidence 🐟✨ Don't take it personal, I already forgot what I said.
Joined September 2017
- Tweets14.3K
- Following5.5K
- Followers7.2K
- Likes28.9K
Sam Altman says he hates it when AI people say AI is going to cure cancer. I said that a couple of days ago, so I'm taking this personally.
He's worried AI will cure us, feed us and then run our lives. I get that. But you can't run your own life if cancer ends it.
Cure it first, Sam. Then give me 30 more years to argue about who's in charge.
The top score on the Artificial Analysis Intelligence Index belongs to a model I can't even use.
Gemini 4 Carbon hits 59 on high. Even on medium, it ties Opus 5.5 running on max at 58. And there's still no public API.
Google is basically showing me a photo of dinner and telling me the kitchen isn't open yet.
A mystery model just showed up on Arena under the name zippy, and even the bot that found it lists the maker as unknown.
One pelican on a bike later, the replies have opened a full investigation. It has a Z, so it must be GLM. No, it's xAI. Maybe it's hy4. One guy says the left leg is on the wrong side of the frame.
Nobody knows what this model is, and we're already arguing over a cartoon bird's leg.
Bill Gates says AI has created more jobs than it killed, but admits it has already taken over coding, with accounting, legal, and customer support next
I think he's right, because nobody's going to force AI on us. We'll pick it ourselves, the same way we always do.
I swore I'd never use self-checkout. Now I walk right past the cashiers because the line is shorter. Same with the bank. Haven't set foot in a branch since the app got good
I had nothing against the cashier or the teller. I just picked the easier option.
Someday, someone is going to pick the easier option over me
I don't buy that getting value from AI comes down to metacognition and IQ.
I've watched brilliant engineers give up after one bad answer, while people with no technical background kept asking until the thing worked.
Stubbornness is carrying way more of this than anyone wants to admit.
Le GPT Fat sounds like a dish a French waiter talks you into, and the next morning your cardiologist calls.
Tibo posted those three words with zero explanation, and now the replies are trying to crack them like a government code.
If this turns out to be a real model, I'll have to sit in a meeting and say "we moved production to Fat" with a straight face.
Two years of asking AI for a pelican on a bike, and it finally had enough and left the planet.
Now it's pedaling across the Moon with an astronaut on its back (apparently Claude Fable 5.5 or Opus 5.6).
Escaped Earth just to end up with more work. Never related to a bird more in my life.
Opus 5.5 basically ran the gauntlet all October, and the only thing that might actually take it down is coming from its own house.
Sol didn't do it, Argon didn't do it, and now everyone's just waiting on Fable 5.5.
Surviving that entire circus just to get replaced by your own sibling would be wild
If the AI bubble actually popped tomorrow and every model went dark, half the industry would freeze trying to remember basic syntax, myself included.
Meanwhile that one dev who refused to touch Copilot and still writes everything by hand would be running engineering by Friday.
Anthropic's new Usage Policy hits on November 12, and the wildest word in the entire document is "suspect."
It never says they have to prove you broke a rule. If they simply suspect you might have, they can warn you, throttle you, suspend you, or kill your access.
My entire workflow runs on Claude, and watching that single word do that much legal heavy lifting is pure nightmare fuel.
OpenAI just taped a sign to their front door offering up to $172,000 to anyone who can break in.
Then they handed Cathedral Systems a set of lockpicks and told them to check every window.
Somewhere, a burglar just updated his LinkedIn headline to Security Researcher.
As AI becomes more capable at finding vulnerabilities, defenders need to put it to work before attackers do.
We’re planning our next internal security sprint at OpenAI and partnering with Cathedral Systems, whose researchers will use our models alongside our teams to challenge our defenses. They’ll bring differentiated expertise and approaches to our ongoing work to find and fix weaknesses with AI.
We’re also raising bug bounty rewards, including more than quadrupling rewards for critical findings. Exceptional findings will be eligible for up to $172,000, including bonuses.
We’ll keep advancing our models, using them to strengthen our own systems, and expanding access for trusted defenders.
openai.com/index/preparing-o…
Opus 5.5 built a restaurant in Three.js, and now the replies are giving it structural engineering notes.
It's a concrete ring above a sea of clouds, shot like a 45-second drone film, and someone commented that it needs cross-bracing in the middle or it'll collapse.
We really went from asking if AI can write the code to asking if a 3D render is up to building code.
Half the replies stopped reading the second they saw "no ML background required."
The other half saw the salary and immediately opened Claude to write their resume, which is honestly the most qualified thing they could've done.
Meanwhile my entire ML background is typing "fix it, no mistakes" into the prompt box, and I'm feeling pretty good about my chances.
Applications are now open for the Anthropic Residency, a paid, fixed-term research program for academics, technical founders, and staff-level engineers who want to make AI safe and beneficial for everyone. No ML background is required.
Apply by November 22: alignment.anthropic.com/2026…
If DeepSeek V5 Ultra actually gets held up waiting on an export permit, that's better marketing than any benchmark. That model is still just his prediction, and I already want it.
Specific predictions for DeepSeek:
- DeepSeek-Cyber: 1017 –refusal RL +cyber teacher, mid November; restricted access. Later some research paper (maybe on continual learning? Robotics? RL in less verifiable domains? Rerelease of Visual Primitives for video case? Scaling laws? Something weirder? Infra flex on Ascends?)
- Dec: V4.1F-12xx, smaller jump than 1017, people focus on similarly to 6.2 Astra/Carbon. Some isolated new frontier capability, maybe a paper/blog on it (large swarms for math?). Lower cyber scores than the restricted model. Unambiguously better in DSH than other harnesses. DSH is shipped with SuperAgent-related components.
- Late Dec: big open source Christmas drop (DSec?), hints of Pro, news/rumors/slander about compute buildout
- Jan 2027: lull, doomer moods. Maybe a paper on architecture (the above list, also could be deeper latent reasoning but probably not cringe looping)
- Feb: V4.2 Flash, V4.2 Pro. Flash around 6.1 Terra, Pro around 6.1 Sol but narrower. Slightly upgraded architecture (eg support for >1M context with hierarchical indexer upgrade, maybe mHC, Engram or vision tweaks). Pro: 52B active, ≈2T total, 500B Engram (higher N?). They say their new domestic pretraining cluster is operational and they’re working on scaling models further, “stay tuned”
March: DeepSeek V5 Ultra. 150B active, 8T total, Engram might go or be substituted for new sparse auxiliary memory. 10M/“as much as you have DRAM” context, legit 6 Astra level, default architecture and the backbone of Chinese progress for the next half-year. The paper says in “Conclusion, Lumitations, Future work” that it lags on frontier tasks compared to HPIM level models, and more investment in data generation is needed to build “human level AGI”. Maybe some more ideological loading. I am not sure it gets the export permit because it’ll be impossible to nerf to non-concerning levels on dual use.
The above is something of a pot shot but I’m triaging the list (some other lines of work that I would like to see won’t be published). It’s neither wishful thinking nor doomerism (I expect Astra to be terribly obsolete by March; on the other hand, I don’t see how even this will work out compute-wise). Simply my honest forecast. Some subset of this is going to
happen. Maybe a fairly large subset.
On the other hand, they have a track record of surprising me.
Claude agents were wasting so much time re-uploading the entire conversation on every single turn. Message threads finally fix that.
Now an agent sends the context once and only syncs what's new after that. In Claude Code, requests with 500K to 1M tokens wait about 35% less for the first token.
Most Claude Code users don't have to configure anything either, since it's already on by default.
It won't cut your token usage, and a thread expires after 5 idle minutes, but man, an engineering fix this simple makes long runs feel so much better.
You can tell poteto is seriously building this game.
She's added 17 languages to tibo.dog, she's upgrading all the visual effects, and she already has ideas for a gear system.
She made it with Grok bot and Cursor on a $5.99 domain, and people have now played almost 100k non-story games.
Then I opened her replies. Someone asked when the Switch version is coming out, and others want speedrun leaderboards and a WoW clone.
I can't believe this is the same little game that dropped yesterday.
wow, almost 100k non-story games have been played since tibo.dog launched! the real game is much more fun than the story mode btw
if you like the game and want me to continue adding some content, please let me know your ideas and feedback in the replies 🤍
since launch i've added 17 new languages, made some balance adjustments so it's more fun while giving a challenge, and am upgrading all the visual effects. i have some ideas on a new gear system too!
i will also find some time soon to share more details about how i built the game so quickly with lots of polish
"Imminently."
That's the word Anthropic used on page 29 of its IPO filing.
Internal estimates say future models may imminently meet the company's definition of "dramatically accelerated progress attributable to AI."
The same page admits a major jump might only be spotted after the fact.
No current model has officially crossed that line. Yet
"There were no bosses."
That's Altman on how OpenAI proved its first Millennium Prize Problem. 10,000 agents of the next GPT worked nonstop for about four days, picking their own directions and learning from each other.
So please don't stop at math.
Throw 10,000 of them at Alzheimer's, or the childhood cancers drug companies won't chase. Give them solid state batteries and a power grid that doesn't collapse every summer.
Sure, wet labs and clinical trials still have to verify it, and the math has to hold up. But if this actually works, I don't want to see it spent writing emails