i’m figuring it out | AI @limitlessFT | Prev. @coinbase
Joined August 2018
- Tweets 4,960
- Following 661
- Followers 56,666
- Likes 60,823
500 Photos and videos
Pinned Tweet
May 15
claude mythos just broke Apple's $2 billion defense system. it did so by discovering a completely different attack vector to break in
only took it 5 days costing ~$35K of mythos api time (the same exploit class costs $5-10M on grey market)
the researchers that commandeered the exploit produced a 55-page report that was delivered to Apple HQ in-person (hoping they release it after patching).
most shocking part for me is apple's MIE worked as intended. mythos just discovered a new way to side-step it entirely by poisoning the data the M5 chip ingested.
at this point i think we have to accept that mythos walks the walk.
As the anthropic red-team explicitly confirmed this week - this is NOT a compute resource issue. its national defense.
❗️🚨 BREAKING: Researchers used Mythos Preview to find the first public macOS kernel memory corruption exploit on Apple's M5 silicon, they give a glimpse into Mythos say it’s really powerful.
Apple spent five years and an estimated several billion dollars building Memory Integrity Enforcement (MIE), the hardware-assisted memory safety system built around ARM's MTE. It was the flagship security feature of the M5 and A19, designed specifically to kill the entire memory corruption bug class.
Researchers from Calif built a working exploit in five days.
According to Apple's own research, MIE disrupts every public exploit chain against modern iOS, including the recently leaked Coruna and Darksword kits. Calif walked into Apple Park this week and handed over the report in person.
Full 55-page technical report drops after Apple patches the vulnerability.
127
457
5,355
1,590,875
23h
amazon owns ~17% of anthropic (currently worth $170B) and supplies them 1 million tranium chips to train/inference claude
this would be utterly insane if true.
Jun 13
Wall Street Journal is reporting that Amazon reported the jailbreaks to the Department of Commerce, who instituted the ban
5
3
80
15,342
Jun 13
my god.
the U.S. government just ordered the shutdown of Claude Fable and Mythos 5 for every foreign national living inside and outside of the U.S.
they cited a jailbreak that gives a user access to restricted capabilities that pose major cybersecurity and biological threats.
anthropic’s response has been to immediately cut off every users access. no one can use Fable right now.
there’s a bunch of issues with this:
- the government is essentially blurring the lines of nationalised and private tech.
- anthropic thinks the jailbreak is actually not a major threat, it’s non-universal meaning they can patch it.
i’m guessing passport/identity checks are coming to claude soon.
Jun 13
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.
The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.
Access to all other Claude models is not affected.
We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.
Read our full statement: anthropic.com/news/fable-myt…
10
1
40
6,647
Jun 12
a peculiar phenomenon - fable 5 speaks in it's own dialect of english when preparing a response
the words are english but they don’t make sense when you read them in a sentence.
anthropic actually called this out in the mythos system card this week:
"mythos is denser and more difficult to interpret… containing more jargon"
it feels almost sci-fi-like. like the ai has eclipsed human's way of communicating
new challenge will be understanding how these models even think as they grow more intelligent
gpt 5.5 and other model labs demonstrate the same trend. they compress language into random words.
imagine an ai being responsible for x% of the worlds GDP and not being able to understand how or why its doing certain actions
Jun 12
Fable feels superhuman at working over long agentic conversations, sometimes to the point where I can't keep up with what it's telling me 😅
This prompt snippet has been the best fix I've found for getting it to write clearly and drop any jargon:
3
22
4,160
Jun 12
SpaceX is officially worth $2.3 trillion (7th largest company on earth)
its entire vision relies on launching 1 million refrigerator-sized GPU racks with 70 meter wing-spans into outer space
assuming elon pulls this off, spacex will create 120GW of ai compute (roughly 3-4x terrestrial data centers)
best part about this is there is no in-between. spaceX either successfully reduces the cost to enter space or it fails
i’m betting on the former. hate him or love him elon is the best 0-1 hardware scaler on earth
book it.
3
9
125
16,336
Jun 12
tomorrow spaceX will raise 3X the largest IPO in history. $75 billion & elon becomes the worlds 1st trillionaire.
the craziest part is i think its undervalued (just hear me out)
- their ai1 satellite is the answer to “will ai datacenters work”. the satellite is a 72 gpu rack, they’re launching 1M of them totalling 120 GW of compute
that’s 4X total ground data centers.
right now the launch costs are more expensive but assuming they bring cost down to $200K per rack it ends up being cost-competitive to data centers on earth
best part is the GPU racks are interchangeable meaning google tpus, amazon tranium and whatever else is welcome. chip-agnostic.
anthropic, cursor and google alone will pay spacex $50-100B over the next year for access
there’s obviously a lot of execution risk but no one scales hardware 0-1 as aggressively as elon.
the core constraints in ai are compute, memory, power and distribution.
spaceX excels at 3/4 of these. it’s an index bet on AI infrastructure.
Teams are go for launch with a $135 price per share for the SpaceX IPO → spacexipo.com/#priceannounce…
8
2
41
5,576
Jun 11
AI is getting way too expensive and its the reason markets might blow up BUT not for lack of demand - its the exact opposite:
> demand for ai is skyrocketing. the models actually work but the tokens are too expensive.
i switched to claude fable 5 this week and hit usage limits almost immediately
who wins? chinese open models. fraction of the size, 80-90% of frontier capability, costs 1/200th of claude fable, gpt 5.5 high.
ignore the benchmarks btw, no one in the real world actually cares, its the bill they have to pay:
> the number of US startups shifting their claude subscriptions to cheaper models has 3X'd recently.
companies like uber are starting to substitute models for the same tasks, allocating cheaper budgets
the bottlenecks are (yeah you guessed it!) compute, power, memory, cooling etc
my guess is the push to cheaper models will force new model architectures and chip design
apple's memory chip design for their new 20B on-device model is a clear example of this.
chart below from latest citadel report:
7
3
29
2,966
Jun 11
i will die on this hill: Apple is a frontier AI company. they will inject AI into every place you scroll, type, read and listen
imagine if you could “google” a past memory or ask siri to prove you’re right in a group chat
ai models that run on-device with access to your entire life.
the craziest part is the models don’t even need to be big!
a 20B frontier model can run smoothly on apple’s chip architecture completely privately
there are 3.5 billion apple devices, the gravitational pull this company has on societal culture is unmatched
14
42
759
135,511
Jun 10
100% - claude fable 5's coding capabilities effectively make humans redundant as a "traditional" software eng
agents now write, check and test code. they even monitor the health of the app for maintenance.
the future sits in these "cloud factories" (mainly the harness) because humans shouldn't be coding at all, they should be orchestrating:
company-level:
> custom software costs pennies and is built in a day. what are you paying humans per hour for???
industry-level:
> if costs drop then smaller players who were priced out (e.g. dry cleaners, schools) now get custom tools for free. this moves the constraint to speed vs. build
regular-person:
> if you're a customer complaining about a problem, your issue gets fixed overnight by an agent.
also the final state of all of this is an AI DOING ALL OF THE ABOVE. why should you be the one to prompt an agent?
anthropic's been ahead of the game here with their dynamic workflow feature.
Jun 10
My take 24 hours after Fable 5:
Your organization will likely not scale with the exponential curve of AI.
I'l just come out to say: This should be a wakeup call for engineering teams.
Set up your cloud software factories. Now.
Models can now fix impossible bugs, UI-test the hardest flows, writing extremely good code, etc. I have't opened Datadog manually as far as I can remember.
AI should be the first-line defense for bugs and feedback. Humans should only look at PRs after an AI has already reviewed it. AI should generate screen recordings of any PR before a human eye even reaches it. The agent should just prompt itself most of the time.
Ex. (pictured) our ui feedback channel manages itself, creates tickets, assigns itself automatically
You might also be worried about cost. Anthropic, OpenAI, and other labs will likely continue to put out bigger and more expensive models. But, we will also continue to get more capable small models. Not everything will need the smartest models. It's about having the organizational harness in place to continue taking advantage of this rising tide.
Moreover, if you use Devin, we've already optimized our harness a bit, and Fable is actually only ~40% more expensive in practice (vs the 2x people assume). I'm honestly pleasantly surprised - it might be higher ROI than you think.
Anyway, if you take anything away, engineers shouldn't be manually picking up tickets, humans shouldn't be digging into logs themselves, rethink what you do with your time that shouldn't just be an AI. We need to rethink what humans spend their time going.
2
2
30
4,981
Jun 10
crazy example of a real malware exploit attackers are using right now in claude fable 5, it involves intentionally triggering fable's safety classifier and sneaking in malicious code
it goes something like this:
-> attacker submits an npm package intentionally titled with a trigger word (e.g. biological weapons design) causing fable's ai scanner to review it
-> scanner sends it to an llm to review it but it reads the title FIRST and immediately marks it as a "safety red flag" and hands it back to the ai scanner
major issue: the attacker included a malicious code injection in the body of the package. but because the llm only read the title (and not the entire piece) it goes undetected.
the result? the ai scanner doesn't flag it as malicious code, just a safety red flag
which means millions of developers around the world can still download it online putting their systems at risk.
seems like anthropic shifting the safety system to the deployment design is creating new attack vectors for hackers
pretty wild
Jun 10
NEW: malware developers added nuclear & biological weapons text to to their spyware.
Goal? To trigger LLM safety refusals... so that their spyware wouldn't be analyzed by an AI security scanner.
Cleanest practical example I can think of for why over-indexing on first order safety alignment is risky.
When closed (and open) models ship with aggressive refusals, they will be sprinkled with second-order blindspots that attackers will discover...and exploit.
We are only in the earliest days of attackers leveraging these features, and it wouldn't surprise me if users systems that need to handle complex cybersecurity issues demand that models be less safety-blunted.
In the weeds: @SocketSecurity's post also shows why intention matters in how you design a malware analysis pipeline to avoid prompt manipulation.
H/T to colleagues that shared this with me socket.dev/blog/mini-shai-hu…
5
7
55
7,823
Jun 10
first 10 hours of using claude fable, mixed reactions - the model's fantastic at most things but unecessarily restrictive in this initial release:
- this models BURNS through tokens, hit my limit within 2 hrs, had to pay for more usage
- its an amazing world simulator. crushes at building game engines, video games, nails visually stunning graphics unlike prior models
- really annoying usage restrictions prevent me from researching basic science blogs. the sandboxing needs to be loosened
- you can't even analyze anthropic's own system card for fable citing cyber security risks.
- creative writing is amazing (as its name suggests)
- much less sycophantic / dumb-sounding vs opus.
- basic math exploration defaults to opus 4.8
im hoping anthropic loosens restrictions asap so the public can really test this models capabilities. think it has huge potential.
8
2
75
12,035
Jun 9
hell yeah anthropic just released 2 of their most powerful models claude fable 5 and claude mythos 5. they're officially #1 and less than half the cost of mythos-preview:
-> claude fable = restricted version of mythos for public use
-> mythos 5 = anthropic's unrestricted model for private use only (government and glasswing partners) takes the crown for #1 model by a wide margin
-> fable's unbelievable at coding. Stripe used it to accelerate a coding project from 2 months to days.
-> vision-mode turns fable into a frontier model that can see everything you can. it beats pokemon fire red on minimal settings and produces visually stunning images
-> cost is only 2x opus 4.8. surprising and way more accessible for mass use
-> drug-design: mythos-5 accelerated parts of the protein design process by 10x (zero human help)
-> fable has strict safe-guards in place, if it suspects you're asking a hazardous prompt it re-routes to opus 4.8
important to note: access is limited to june 22nd at which it moves to a usage-based model due to limited compute.
Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use.
Its capabilities exceed those of any model we’ve ever made generally available.
5
2
30
3,904
Jun 9
apple’s siri AI looks dumb until you realise it’s about to launch on 3.5 billion devices overnight, tap into the richest data set in consumer history and easily improve the lives of its users in some obvious ways:
-> siri’s always on (even offline), works across any app.
-> trained on your texts, calendar history, email etc = instantly the best personal agent
-> apple not charging for use (unless usage limit result?)
-> models run privately on-device via google cloud.
best part is apple can use any ai model. they could easily partner with anthropic or openai to serve their models, they’re agnostic - but in doing so siri ends up being the orchestrator.
genius patience from apple to chase the model hype and focus on distro.
Jun 8
Just got into Siri AI ... asked 'when was my flight last month?' and it knew. Good start.
16
11
374
51,168
Jun 8
google exercising raw power of being top dog by offering frontier intelligence cheaper than any other model provider
goal: drive adoption by being the cheapest provider whilst starving model labs who can’t offer the same rates for as long
you may not care about cost but enterprises burning $10-100M per year def do.
only other companies that could do this are apple (who’s using gemini) and meta (who’s model kind of sucks rn)
Jun 8
📣We're updating the price of our Google AI Plus plan to $4.99/mo💰or local equivalent (down from $7.99), and doubling the included storage, from 200GB to 400GB ☁️. Now you can unlock tools to boost your productivity and creativity - and get more space to store your photos, videos and projects - for less.
2
1
30
5,381
Jun 8
once again a kid with no investment experience is out-performing the best investors in the world
leopold's latest investment holdings show a massive $4 billion stake in anthropic (20% of the fund) and 270% returns from this year so far. his aum is now $20B
looked into it and he invested in anthropic at a $60B val back in feb 2025. anthropic is now worth $965B, 16x on 1 position.
for context, bill ackman's fund pershing capital is around the same size and they've been around 22 years
don't think we've ever seen someone grow capital this quickly in a fund
SITUATION DETECTED: Leopold Aschenbrenner’s Situational Awareness is up 270% after fees this year and more than 1,000% since inception, per WSJ.
The fund now has $20B under management and counts Jane Street as an investor.
25
38
854
296,353
Jun 7
tomorrow Apple (finally) becomes a serious AI company.
Siri will become the #1 ai agent simply by sitting on top of other frontier models and orchestrating prompts to the best one
they’re 3.5 billion apple devices globally, there is no company in the world that knows consumer intent than apple.
64
12
156
46,158
Jun 7
god these open models are actually such good value for money vs frontier ai
this thing is 18x cheaper than opus 4.8. it’s not as smart but it probably does 50% of low-level automation easily.
i assumed open models always lagging frontier was a bad thing but i guess companies won’t care if it does the job
they don’t need frontier for everything
lots of american silicon valley startups using chinese models. more open source models popping up in agent harnesses. nvidia launching nemotron 3…
honestly never thought i’d say it but open source could have a permanent home ok the business world
We gave the same code audit to Claude Opus 4.8 and MiniMax M3.
Same codebase. Same prompt. 17 known bugs planted in advance.
MiniMax M3 caught 13 of them for $0.07. The cheapest Claude run caught the same 13 for $1.30.
Here's the breakdown. 🧵
ALT Black-and-yellow comparison graphic from Kilo Code’s Code Audit Benchmark. The headline reads “MiniMax M3 vs Claude Opus 4.8.” Below, benchmark results show 17 bugs planted, with both models catching 13 bugs. Cost per run is highlighted as $0.07 for MiniMax M3 and $1.30 for Claude Opus 4.8. A note at the bottom reads, “Same codebase. Same prompt. Full breakdown in thread.” The Kilo Code logo appears in the upper-right corner.
5
3
30
4,614
Jun 6
ai’s ultimate form will be an always-on assistant that lives across every device, messenger app and work platform you use - kinda like a digital shadow
you can talk to it using voice, it instinctively recalls memories you forget, it spawns multiple agents to do your bidding
always on and always consuming tokens night or day
this is also the biggest reason why picks and shovels companies will always be in greater demand
literally no one even uses these tools yet
Jun 6
wait so this turns ai into real-life JARVIS from ironman. thats so damn cool
100% of my interaction with LLMs has been voice to text in the last 3 months (exc some code)
looks like this cursor update now lets you gesture at things on-screen and the ai intuitively knows what you’re referring to
idc what you say that’s a MUCH better experience than screenshotting random shit
13
2
18
3,417
Jun 6
wait so this turns ai into real-life JARVIS from ironman. thats so damn cool
100% of my interaction with LLMs has been voice to text in the last 3 months (exc some code)
looks like this cursor update now lets you gesture at things on-screen and the ai intuitively knows what you’re referring to
idc what you say that’s a MUCH better experience than screenshotting random shit
Jun 5
Working with agents should feel like working with a colleague. You should be able “speak to” them not just with text chats, but by gesturing at a screen together, talking live, etc.
3
3
18
7,359
Jun 6
as a former biologist that spent his days sequencing genes this is absolutely wild to see
claude can now do highly specialized work that scientists typically needed custom tooling for, but thats not even the shocking part:
-> the insane thing is this just regular opus 4.7! not even a specialized fine-tuned version. generalized llms are now good enough for frontier science tasks
-> this will flatten the barrier to scientific discovery dramatically over the next year
-> anthropic, openai and google each signed a government petition this week calling for immediate regulation of ai bio models warning of ease to create bio-weapons.
am i the only one that thinks science is the next domino that falls after coding?
Jun 5
New Anthropic Science Blog: Making Claude a chemist.
To manipulate a molecule, chemists first need to understand its structure. Their main tool is NMR spectroscopy.
We found Opus 4.7 matches—and on some tasks beats—dedicated NMR software. Read more: anthropic.com/research/makin…
12
16
158
16,565
Jun 5
elon's on a generational run right now, google just agreed to pay spaceX ~$1 billion per MONTH for 3 years
-> google accesses 110,000 gpus, cpus and memory in order to scale Gemini
thats another $30 billion revenue booked on spaceX's balance sheet bringing the total to $85 billion across anthropic, cursor and now google
3 partners!
IPO expected in 1-2 weeks.
May 12
it looks like an AI alliance is gradually forming between SpaceXAI, anthropic, tesla, google and cursor.
its incredibly beneficial to all parties involved:
> google gets cheap access to space infinite energy
> anthropic gets 300MW of inference compute
> spaceX gets $5-10B between anthropic and cursor deals
> cursor gets a leading coding model from the compute
both google and anthropic have now publicly announced intentions to use SpaceX to launch ai data centers into space in the last week
Google owns 7% of spaceX and 14% of anthropic so it makes sense
6
4
69
7,365