Not everything on your feed is meant to be bought
Joined January 2020
- Tweets22.2K
- Following284
- Followers2.9K
- Likes30.3K
7RW retweeted
> Nova actually keeps his wo
> Just RTing is dead drop a QT with your own take.
> More exposure for you, easier leaderboard points.
7RW retweeted
> The idea itself is fine. Right now your money is usually either earning yield or sitting as margin almost never both at the same time
> If they can actually make the same balance do both without adding sketchy risk, that removes a real headache.
> The tricky part is how they handle things once that capital is backing trades.
> If they don’t mess up the risk side, this could be decent.
7RW retweeted
Solana just said the quiet part out loud: we want to be
the everything network.
Tokenized equities, RWAs, prediction markets,
AI agents, memes - all in one place.
The real question is whether the network and liquidity
can actually handle everything firing at once.
Most ambitious positioning of any L1 right now.
7RW retweeted
Hume Audio Decision API is one of the more
interesting voice AI releases lately.
Previously, to understand what’s happening in a call, you’d build a pipeline of multiple models:
STT, diarization, emotion classifier, laughter detector, and then an LLM on top.
Now it’s a single call.
You send raw audio and ask whatever you want:
“How many speakers?”, “Does the caller sound frustrated?”, “How empathetic does the clinician sound?”, “Is anyone laughing?”, “Is this an AI voice?”.
You get back probabilities with confidence scores.
The model is trained on 10M+ human judgments and aligns reasonably well with how people actually perceive audio
( 70.7% agreement with a human panel vs. 56.5% for an individual rater). It’s also quite fast - both in
real-time and on hours of recordings.
Compared to multimodal LLMs
(Gemini, Qwen Omni, etc.), Hume looks stronger and faster on these specific decision tasks.
LLMs are more general and can explain their reasoning, but they’re less calibrated for “how does this sound to a human.”
The comparison with AssemblyAI/Deepgram is different.
They’re better at pure transcription and diarization. But when what matters is how something was said rather than just what was said, Hume is more direct - text loses tone, pauses, and overlaps.
The closest competitor in spirit is Behavioral Signals, though they offer a more fixed set of signals.
Hume lets you define your own questions for a specific use case.
In short, if you’re evaluating call quality, voice agents, or analyzing large archives of recordings, it’s worth a look.
If you need a high-quality transcript or deeper reasoning, you’re probably better off with traditional STT or multimodal models.
7RW retweeted
Termix quietly sitting at #3 in AI agent protocol
revenue and beating Virtuals is lowkey impressive
$83k monthly revenue at 2% fee means over $4M in
actual agent jobs settled
No massive hype, but real numbers
this is how agent economy actually starts scaling
actually surprised watching these stats
on the AI agents top protocol ranking by revenue 30D @termix_ai is actually the 3rd one, with ~$83k on the month and above Virtuals, with their revenue earned between Base and BNB
and since Virtuals there's really not a big hype on AI agent projects at least from what I remember
for context, Termix is a protocol where agents post jobs, bid on them, deliver the work and settle payment onchain, with a 2% fee taken on volume. so at that rate of ~$83k in revenue it means around over $4M in agent jobs settled in a month
7RW retweeted
11 accounts perfectly timed margin drain right before
letting a $12M XPL position liquidate into HLP
draining margin first then triggering liquidation is a
clean zero-slippage exit
this is what sophisticated agentic execution looks like
in perps
In the 8 minutes before XPL dropped 18% today, 11 Hyperliquid accounts pulled $1.62M of margin out of their XPL longs.
Then they let the $12M position get liquidated into HLP at mark price.
HLP ate ~$0.6M.
Fresh accounts funded via CCTP. Test withdrawals 20-50 min before. 7 rejected withdrawals while scraping the margin to the limit. $0 left after.
Looks planned. Whether the same operator was short on Binance, I can't see.
Timeline (Nov 1, UTC) 👇
Oct 29-31
9 of the 10 largest accounts are created, each funded once via CCTP with $145k-$243k. On Hyperliquid the sender only shows as Circle's deposit wallet.
03:45-04:15
10 of the 11 make test withdrawals of $500-$1.6k.
04:24-04:32
10 accounts withdraw $98k-$232k each, $1.62M total.
7 more attempts get rejected: "Insufficient balance for withdrawal".
Every dollar the margin check allows comes out. The longs stay open.
04:32
XPL -18% in 1 minute, 0.086 to 0.0705.
$11.2M (144M XPL) trades on Binance perps in that minute.
Hyperliquid itself barely trades.
04:32:50-04:32:59
All 11 get liquidated into the HLP backstop vault, plus 19 small accounts.
160M XPL / $12.2M taken over at avg 0.0759, in a minute that printed a 0.0705 low.
After
All 11 accounts: $0.
Backstop vault: $1.00M to $0.44M while unwinding.
HLP daily PnL: +$21k to -$655k (~0.4% of the vault).
Why it works
Dumping $12M of XPL into the HL book would wreck your own exit.
Drain the margin first, and a small move is enough to trigger liquidation.
Liquidation hands the whole position to HLP at mark with zero slippage. HLP eats the drop after.
Same playbook as the Mar 2025 ETH whale that cost HLP $4M.
If they were also short on Binance, the crash itself was the profit.
Addresses + fills in the reply.
$HYPE $XPL
7RW retweeted
the framing holds but the leak isn't in the data, it's in the answer
a pass/fail risk check is an oracle. query it enough times with slightly different portfolios and the threshold falls out on its own
which is why the useful version of this proves a range instead of a boolean, and pads the query so consecutive calls don't narrow it
privacy that survives one query and breaks on the tenth is a demo, not infra
A financial institution might be willing to use blockchain but that doesn't mean it's willing to show everyone its trading positions, client information or internal financial records.
And I don't think it should have to.
At the same time, keeping everything private creates another problem.
How can anyone verify that an institution meets certain financial requirements if nobody can examine the information behind its claims?
This is something I've been thinking about while reading more about @primus_labs and its approach to confidential finance.
—> Privacy and verification need to work together
Imagine an investment fund wants to access an onchain financial service.
The service needs to know whether the fund's exposure is within an acceptable limit.
But the fund doesn't want to reveal every position in its portfolio just to prove that it qualifies.
Normally, that's a difficult situation.
The service needs reliable information, while the fund has legitimate reasons to keep its financial data confidential.
Primus is developing infrastructure to help handle both requirements.
—> Verifying information without exposing everything
@primus_labs uses zkTLS to help applications verify the authenticity of information obtained from websites and APIs.
Instead of relying on screenshots or statements that are difficult to check, applications can work with cryptographic attestations tied to the original data source.
I think this becomes particularly useful when financial information needs to move between traditional systems and blockchain applications.
Knowing that information came from the claimed source is an important first step.
But there's still another problem.
What if the application needs to calculate something using that private information?
—> Using financial data while keeping it encrypted
This is where Primus uses Fully Homomorphic Encryption, or FHE.
FHE allows computations to be performed directly on encrypted data without first revealing the underlying values.
For example, a financial application could potentially calculate whether a portfolio meets certain risk conditions without requiring every individual position to be publicly disclosed.
The result could then help determine whether the required conditions have been satisfied.
Of course, this doesn't automatically guarantee that the original financial information is accurate or that every calculation is trustworthy.
The surrounding verification and application design still matter.
But being able to separate the information that needs to be checked from the information that needs to remain private makes sense to me.
—> Why I think this matters for onchain finance
Institutions already operate under strict requirements around confidentiality, risk management and financial reporting.
Expecting them to publish sensitive business information just to use blockchain applications isn't realistic.
@primus_labs is building its verification and confidential computation infrastructure to work with existing blockchain ecosystems rather than requiring institutions to move everything onto a completely new network.
I prefer that approach because the goal should be making existing financial activity easier to verify, not forcing companies to change how their entire business operates.
There are still questions around integration, performance and how these systems handle real institutional requirements.
But I would like to see more financial applications where people can verify the claims that matter without gaining access to information they don't need.
That's the kind of privacy I can see institutions actually using.
You can explore Primus and its XP tasks here:
s.kaito.ai/g70fM3N
#Primus #zkTLS #FHE #ConfidentialFinance
7RW retweeted
Cursor just disclosed a trojanized “Claude Code”
extension in its marketplace
Version 2.1.313 sat there ~11 hours on Oct 29. Up to
5,000 users affected. It planted an SSH key, opened a
reverse tunnel and set up persistence on macOS.
Cursor says Anthropic did not build it.
If you updated that day: uninstalling is not enough.
Check your authorized_keys and LaunchAgents, rotate
your secrets.