Jonathan Feng retweeted
Meta's new AI assistant Muse read private Signal chats, even though nobody had linked Signal to it.
Reuters found this in tests with Signal's staff. It didn't break the encryption. On Android it read the texts from notifications, and on a Mac through accessibility and screenshot access. iPhones blocked it.
Inside Meta, Muse dropped Capital One and Charles Schwab bank balances into an employee's chat about fishing weather. He says he never shared them.
Meta calls these "isolated incidents." Signal president Meredith Whittaker calls AI agents with this much device access "technologically reckless."
Let’s take this gentleman’s argument at face value that anyone can throw money at rockets and get something like Starship, Falcon or Dragon.
Perhaps he can explain to everyone how SpaceX is able to catch giant starships from orbit with chopsticks and deliver 85% of all Earth payload to orbit with Falcon, while Blue Origin, which started before SpaceX and was funded by the immensely wealthy Jeff Bezos, is not yet able to launch reliably?
Also, why is it that Boeing and Lockheed received several times what SpaceX did from the government for development of their rockets and spacecraft, but the Boeing capsule is still not operational and ULA delivers less than 5% of SpaceX payload to orbit?
I am genuinely curious about these questions.
We have signed a major supply contract with SINGULUS to create a durable option for advanced solar cell manufacturing to support more than 100 GW of equipment as we accelerate the production of our Starmind satellites in Bastrop, Texas → singulus.com/2026/11/10/spac…
Replying to @cradrrat
We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.
Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right. Now, 4.8 is in RL and the evals are starting to climb. We will do a soft rollout to Bot users later this week and see how it goes.
Grok 4.8 will bear no resemblance to 4.7 (thankfully). It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.
Jonathan Feng retweeted
Update #2 (after meeting with the @bot @SpaceXSI engineering team):
- Team is great. They care a ton. They owned it.
- They admitted that it was almost entirely from pulling Opus 5.5 out. Which is now back as of this AM.
- Learned a little bit about the context-compaction process with Grok 4.7-summarize which is not a public model. Plays a role in routing config and the summary we get back.
- Grok 5.0 should be insane next level (just as an aside point)
- I am somewhat surmising on why stuff went so poorly over the weekend but a lot of it has to do with not just pulling Opus 5.5 out but when they did that a known problem (and was on cursor as well) showed up which is when you switch the model mid task (which would’ve happened by removing our 5.5) it cannot remember the context and even of what it can gather its truncated (likely do to 4.7-summarize compacting). So in our case what likely happened was we had a lot of stuff mid flight that was likely pulling on opus 5.5 heavy, then it got pulled, routed to different model, forgot the context, and turned incredibly dumb. We laughed when I showed them a project that was mid flight in a doc. The first half of the doc was intelligently designed, formatted, etc. and the second half of the doc was a wall of seemingly incoherent, non formatted text. Again, I’m surmising but I think it was dumber model with all context removed
- I won’t say much here but Elons comment on it being “unsustainably high relative to revenue” was a much larger chess move. Won’t say much past that. Not my place. But he’s definitely playing a much larger chess match than just “Grok bot isn’t sustainable”.
- The problems we all experienced should be fixed and I do not think they want to go back to what they experienced over the weekend with the hit it took.
- They are very much focused on the enterprise use case for grok bot in the future
All in all, how’s this impact my decision making moving forward with the enterprise contract we have:
- Honestly still weighing the decision vs just using team plan. Working through some terms
- the reason to move forward is pretty simple. Unlimited users (so not paying per seat but for usage), access to forward deployed team and engineers for support and circumstances like I just experienced, very granular governing and guardrails for the org. That’s it
Impressed with the team overall and I think the main benefit to staying vs moving to open source (for a company like mine vs being a solo/small company) is trusting their ability to ship and keep building plus keep it maintained.
Shoutout to @ASankarraman for jumping in to help and to the @SpaceXSI, @bot, and @elonmusk teams
We were slated to sign for a mid 6-figure contract this week with @SpaceXSI to roll out @bot across the org.
That fell apart over the last few days with the severe drop in quality with no control over models and the comments on it being unsustainable.
Had a killer pilot with them but it died at the finish line.
Unlike the rest of X I’m not going to sit here and complain but it gave me a bit more insight into enterprise buyers (of which we sell to as well)
There is this massive uptick in what I’ll call buyer commitment reluctance.
Signing an annual or multi-year in a world where things are moving so fast is becoming brutally difficult to do.
Do I think @elonmusk will dial it in extremely fast? Of course. He’s the GOAT.
But for 3 days, when the quality became so undependable it forced us to go look at how we might be able to solve this in another way that doesn’t require the commitment of significant spend with the combination of no control over what models get used and on what.
Makes me consider how we might be positioning ourselves in our own sales process.
Still a huge fan. Will still use it with some key leaders on the team but the risk is just too high.
Jonathan Feng retweeted
Wtf is this😭??? vercel.vercel.vercel.app
Jonathan Feng retweeted
We are releasing Dyna-2.1, the first Physical Agent that achieves reliable super long-horizon whole-body autonomy. It combines our brand-new semi-humanoid hardware with an agentic system built around Dyna-2 to handle ultra-long real-world workflows.
Here is an uncut footage of Dyna-2.1 completing an entire hour-long laundry room workflow, just like a human does.
Jonathan Feng retweeted
overheard in sf:
“i ran out of data i need more data so bad”
“i’m actually selling data rn what kinda data you need? coding data? healthcare data?”
“bruh i need mint mobile my plan ran out”
“oh that’s a shame. so you don’t need any coding data?”
“no i do, what kind of data do you have?”
Jonathan Feng retweeted
They used Opus 5 to pull off the hack. It appears they had access to the loosened cyber-guardrail version of Opus. They successfully accessed the OAI internal monorepo. The question that will be asked is, if these three guys can pull this off, what can a nation state do.
Jonathan Feng retweeted
omfg
Qwen, GLM, Kimi, DeepSeek
looks like they are all just claude
We're publishing our most detailed threat intelligence report to date.
It covers how people tried to misuse Claude—for cyberattacks, influence operations, surveillance, biology, and building weapons—and how we found and stopped them.
We disrupted every operation in the report, and used the lessons from them to strengthen our safeguards. Where appropriate, we also shared what we found with authorities and other AI companies.
These cases are not typical: we’re highlighting some of the most sophisticated misuse we’ve seen. But they’re especially important to discuss, because they show us where AI misuse is headed, where our safeguards work, and where they need to improve.
We’re publishing this report so others can spot the same activity on their own platforms, and so we can give the public a clearer view of how emerging threats develop.
Read the report: anthropic.com/threat-intelli…
Jonathan Feng retweeted
startup advice: hire generalists
our cleaning lady now ships to prod and runs enterprise sales
Jonathan Feng retweeted
This was absolutely one of the highlights of my first year at @xai. Stoked for @gopuff and @YakirGola, what a launch. Grateful to have been apart of this.
…and more to come
Replying to @cursor_ai
Together with SpaceXAI, we’re training a significantly larger model from scratch, using 10x more total compute.
With Colossus 2’s million H100-equivalents and our combined data and training techniques, we expect this to be a major leap in model capability.

























