Keeping the human in the loop
Joined February 2022
- Tweets12.4K
- Following218
- Followers351
- Likes6.6K
OpenAI's Control Room stops model output within 35 seconds of a decision to intervene
Here's why it works... the kill switch doesn't trust the workload
The job asks Control Room for a job_id. Every time it samples tokens, the inference engine checks that ID before answering. Stopped job? The engine refuses.
The job code can keep running and it still gets nothing.
If i were building agents, i'd copy that:
- put the permission check in the layer that produces output
- give every run an ID you can revoke
- don't rely on the agent's own code to stop itself
A stop button the workload controls isn't a stop button
Full article here: openai.com/index/control-roo…
9
Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort on FrontierCode 2.0, for roughly half the cost per task
That's from Anthropic's early testing, so treat it as a lead, not a guarantee
But it points at something practical: the effort dial can matter as much as the model name
If i were you, i'd run hard tasks like this:
- start with Opus 5.5, as Anthropic suggests for most coding
- move up to Fable 5.2 only when it stalls
- compare cost per finished task, not price per million tokens
Fable is $8/$40 per million tokens, Opus 5.5 is $4/$20
The higher sticker price can still be the cheaper run
Which tasks have been worth paying for a top-tier model over Opus-class?
Full article here: anthropic.com/claude-fable-a…
1
1
40
First day ever with over 100k impressions on X, and I don't know how to feel about this.
No idea if it's a lot or not, but the new algo feels good, and apparently the trick is to be a legit human being in an ocean of AI chatboxes
1
16
Mira Murati, former CTO at @OpenAI, is talking about moving from turn-based interaction to time-based interactions for models.
Ashamed to say that I didn't know about her till this video, but I feel a similar struggle, especially with AI used in the medical field and with the elderly, who tend to need an interaction as warm and human as possible
49
Nope, wrong, and responsible to sell the dream in this market, especially in this market. And mostly to a younger generation.
What is even crypto ? Are memecoins ? Btc ? Web3 gaming ?
We are a industry that never evolved more than the 30 th perp trading platform.
It s iresponsible to sell the dream when pump fun it's extracting everything out of most ecosystem, with btc under 80k and with global unrest everywhere.
And what skill can you learn in 1-2 months as to have this crazy return in crypto ?
Trading ? Show me a guy who is a profitable trader with 2 months of training.
Marketing and copywriting ? Solved by AI
Community building ? Wat communities ?
And the list goes on
The bottom line is. Don't sell the dream.
It’s crazy how much easier it is to make money in crypto compared to almost any other industry.
And the funny thing is, this has been the case for years. Crypto has never really been hard mode.
Spend 1 to 2 months learning a skill, connections and actually trying to solve problems, and your chances of making money are surprisingly high.
Very few industries offer this level of opportunity with such low barriers to entry.
Take advantage of it while you can.
This made my morning, you put in all that effort as a man to find love, and in the end it s a Chinese spy. Curses
61
You try your best to find love, and in the end it s another communist spy. Curses
The story keeps getting crazier. To recap everything that has been reported so far:
- The US intelligence community discovered that Jensen Huang, the CEO of the largest company in the world, was sleeping with Alice Huang, a Chinese spy and an executive at Megaspeed. (FT)
- Multiple independent investigations have concluded that Megaspeed is likely a Chinese cutout, which helps Chinese companies access AI chips or circumvent export controls. (NYT and Bloomberg)
- The US intelligence community also concluded that Alice Huang’s goal was to convince Jensen to lobby the U.S. government to loosen export controls on advanced AI chips. (NYT)
- The intelligence community relayed this information to “a small group of top US government officials” earlier this year. (FT)
- Over the last year, Jensen Huang has repeatedly lobbied the U.S. government to loosen export controls on advanced chips. @BISgov has done so, not only by permitting the export of H200 chips to China, but also not enforcing export controls that would have required a license for exports of the most advanced Blackwell chips to Megaspeed and Southeast Asia.
- Although @CommerceGov leadership was presumably briefed on this information by the U.S. intelligence community, @BISgov has taken no actions to tighten controls on Megaspeed or Southeast Asia.
- Megaspeed remains a preferred cloud partner of Nvidia, and continues to receive Nvidia Blackwell chips despite employing a Chinese spy who reportedly is romantically engaged with the CEO of Nvidia.
- Tens of billions of dollars worth of AI chips have been exported to Southeast Asia this year alone, and Malaysia and Singapore are now two of the four largest importers of AI chips in the world.
What is going on here??? We need answers. The question of who gets the most advanced AI chips in the world is literally the defining strategic question of our time. And not only does it seem like the MSS may be directly influencing the debate on export controls in the United States, but also @CommerceGov is refusing to listen to the warnings of the IC on this issue. What is a POSSIBLE good explanation for why Megaspeed is still receiving chips? Why is there not a license requirement for AI chip shipments to all of Southeast Asia? This is beyond strange at this point.
The fallout from this story will be massive. Obviously, the motivations underpinning Nvidia’s policy views opposing export controls must now all be questioned. But also, how did @BISgov sit on this information and has still taken no actions? We need answers.
30
Once again, terrible day to be a video editor
Introducing Higgsfield Miracle.
Our latest proprietary post-production technology vectorizes any video into layers for localized, pixel-level editing in real time, with world knowledge built in. Every frame holds the image depth needed for cinema-grade post-production.
Now in private beta with select testers. Coming soon to Higgsfield MCP.
Set up Higgsfield MCP in Claude to be the first to try it.
1
1
45
Greg Brockman from OpenAi, talking about the unreal evolution of the entire medical system as we know it.
And I tend to agree, we are staying too much in our own bubble and we fail to perceive the insane amount of effectiveness AI is showing in the medical field and in all the sub field.
57
Elon Musk just revealed that SpaceX will begin launching StarMind SI computers into orbit in Q3 2027.
You are not bullish enough.
50
Machina Labs got $9M and 18 months from the Pentagon to build metal LUCAS airframes in quickly assembled modular factories
Here's why it's worth watching... it's a trial, not a mass production order
The idea being tested is that a factory should reprogram as fast as the product changes
If that holds, the same robotic capacity could move across drones, missiles and aircraft repair
The real play isn't one cheap drone design
It's whether software-defined production can be reused when the design changes
That's the part to track, not the contract size
Full article here: runtimewire.com/article/mach…
58
OpenAI went in asking for $30B at $1.4T. The first close is $11B at $1.25T
That's the detail that matters, not the headline number
Bloomberg says the round could still reach as much as $20B in the coming weeks. So the target got cut, not the interest
Here's how i'd read it:
- private money still funds frontier labs
- but it wants a lower price and a smaller check
- public-market skepticism on AI infrastructure spend is now part of the negotiation
If you build on these models, watch how the rest of the round fills...
That's your live read on how long the spending race can run
Full article here: runtimewire.com/article/open…
39
Not all jailbreaks are equal, and the Frontier Model Forum just put a rating system behind that
Version 1.0 of its cyber jailbreak framework scores each successful bypass on four things:
- capability gain
- capability scope
- ease of deployment
- ease of discovery
Here's why it matters: without a shared method, two teams rate the same jailbreak differently and burn time patching low-risk stuff first
The real play is to separate "the model said something edgy" from "this gives an attacker real cyber capability, and it's easy to use"
It's preliminary, and they're testing whether different assessors land on similar ratings
Public comments are open through December 21, 2026
Would a shared jailbreak severity score change how you prioritize red-team findings?
Full article here: frontiermodelforum.org/updat…
37
Hey @grok, you have 7 days to teach a total beginner to spot AI-generated images, video and text with confidence. What do you teach each day, and which 3 tricks fail most often today?
1
1
21
SoftBank paid for its $64.6 billion OpenAI stake with a bridge loan, Nvidia share sales and bonds
Now it's reportedly in talks to add several hundred million more. Talks only, no signed deal
Here's why the small number matters less than the funding trail:
- $40 billion bridge loan, since repaid
- $5.8 billion raised selling Nvidia shares
- $11.1 billion in high-yield bonds
The bridge loan is gone, but the bonds still carry interest
OpenAI's valuation went from roughly $300 billion to about $852 billion post-money, so the bar for performance is very high
The real play for anyone watching: track the financing, not the headline check size
Full article here: cryptobriefing.com/softbank-…
49
Hey @grok, which AI predictions from 2015 to 2020 turned out right, and which were badly wrong? Show who made each one and what actually happened.
1
21
If you run agents locally, Ollama v0.40.15 has a quieter change than the new models
Recurrent and vision GGUF models can now serve more than one request at a time
Here's why it matters: a single-request model is fine for chat, but it's a bottleneck the moment you run parallel agents or sub-tasks
Also in this release:
- mimo-v2.6, a 9B model for code and agent work
- ollama launch now supports Kimi Code CLI and offers to migrate old kimi-cli setups
- Pi's model list fills in automatically
If i were you... i'd re-run my concurrent agent setup on a local model before assuming it needs a hosted API
The boring serving fixes are the ones that decide if local is usable
Full article here: github.com/ollama/ollama/rel…
Trump says he's 100% in favor of "Super Intelligence" and will work with states that want to get RICH from it
Here's why the detail matters:
The pitch to states is money, not safety or capability
And it comes with a naming push: an executive order tells federal agencies to say "Super Intelligence" and "SI" instead of AI in official documents
So the federal line is simple: back it, rebrand it, and call the "no" states foolish
If i were you, i'd watch which states actually take the offer
That's where the real rules and incentives will show up, not in the post
Which states do you expect to take the federal offer first, and why?
Full article here: aa.com.tr/en/americas/trump-…
28