Applied AI & research | models, agents & cybersec | still stacking sats
Won’t trap me
Joined November 2013
- Tweets52.2K
- Following1.3K
- Followers2.9K
- Likes224K
while United is restricting ai agents from booking flights, a16z is arguing businesses will soon have to compete for those agents’ money
a16z calls it b2a, business to agent, where your ai increasingly chooses and uses services to get things done for you. for example, they asked chatgpt to make a video ad, and it coordinated across krea, descript and elevenlabs, even recommending a paid descript plan to finish the job, so one request connected several products without someone manually managing every step
blocking agents from accessing your business starts looking shortsighted at this point, because if my agent can save me time (and money) by getting the job done elsewhere, the company blocking it might just lose my business too
The primary users of the Internet are no longer human. They’re AI agents.
As agents move from answering questions to completing tasks, they’ll increasingly decide which products to use, which services to subscribe to, and where to spend money on our behalf.
a16z's Olivia Moore and Justine Moore on the B2A economy: a16z.news/p/welcome-to-the-b…
57
don’t put off this Chrome update, a malicious webpage could run attacker-controlled code inside your browser’s sandbox JUST BY LOADING
Google confirms an exploit for CVE-2026-125350 already exists in the wild and released a fix on november 9 in chrome 157.0.8103.50/.51 for windows and mac, and 157.0.8103.50 for linux
i’d check for updates and relaunch now, because a downloaded update only protects you once chrome restarts
110
could ai research agents push Anthropic’s zk-proof generation beyond a 23x speedup? better gpu kernels alone have an optimistic estimated ~23x ceiling, but redesigning the proof protocol could potentially take it further!
good research on this from a former paradigm researcher. Anthropic’s flashprove uses zk-proofs to verify that a response came from a specific model without revealing its weights, so i (and you) could check that i actually got the model i paid for. generating those proofs is expensive, so it’s useful to analyze how much faster the existing design could get
and the results are pretty encouraging, with an optimistic estimate of up to ~23x faster proof generation from better gpu kernels alone, while keeping the same protocol and its roughly 2.9 billion hashing operations per 1,024-token request in the tested setup
but if ai research agents, which are getting stronger and stronger, help design proof protocols that need less work in the first place, the speedup could go even further from here, potentially making such checks more practical at scale across more models
New post: How much faster could FlashProve get from better kernels alone? 🕵️♂️
I think a defensible hard upper bound is ~2800 tok/s/gpu, but likely much lower in practice. Past that, the next OOM will likely have to come from optimizing the protocol itself.
xn--2-kmb.com/how-fast-can-f…
1
2
177
doubt your ai model can find the optimal path in under an hour on sokobanteg frontier
it’s a fun benchmark for ai models built on a sokoban game. push crates, dodge goblins, and be careful not to brick the room (!)
this is quite a challenging task, need to plan dozens to hundreds of moves ahead. every move (key) is one of 4 directions, so even a 50-key route is one out of about 10^30 possible sequences, so need to chew through millions of positions to find the good one
so it’s basically a shortest path problem over game states, the kind you solve with BFS algo or A*, like solving a 15-puzzle or a rubik’s cube in the fewest moves, just on a map that changes with every push. so it tests long-horizon planning, spatial reasoning and backtracking
a key is just one button press, so the score is how many presses your route takes. a room counts as proved when there’s a math proof that nobody can do it in fewer keys, so the best route and the floor are the same number (just 3 unproven rooms left!)
come join the sokobanteg frontier challenge. it has almost reached my own holdout result (dotted line). the real fun begins there!
soko.banteg.xyz/frontier
1
1,037
“The demand is so high that even if the watt-to-dollar conversion is low, you're going to still use the asset.”
ai demand is keeping older GPUs worth renting, while, as you can see, H200 and B200 rental rates have actually climbed above where they started
GPU rental rates aren't depreciating on schedule.
As of October 19, rental rates were above a six-year straight-line depreciation path from their first observed rent:
- A100s by 9%
- H100s by 54%
@mavolpi: "The demand is so high that even if the watt-to-dollar conversion is low, you're going to still use the asset"
64
turns out Initia got exploited twice in just a couple of weeks, with another $3M+ stolen. is getting hacked multiple times becoming the new standard for crypto projects?
Based on our investigation so far, on November 9th an attacker gained unauthorized access to Initia’s infrastructure through a command injection attack with a compromised GCP key used for provisioning private nodes. This exposed environment variables, which allowed the attacker to access our GitHub and GCP accounts and obtain further credentials.
With these credentials the attacker gained access to Anvil, the hosting platform used by rollups on Initia. Through this the attacker was able to validate forged IBC withdrawals from five rollups, unlocking assets held by the Initia L1’s IBC escrow accounts.
Around $3.27m of assets were moved to the attacker’s Initia L1 address with $2.16m of the stolen funds remaining on the paused L1 where they cannot be moved. The attacker sold stolen assets on the Initia DEX and then withdrew 13.4k TIA, 8k USDC.inj, and 597k Noble USDC reaching Initia’s outbound IBC rate limits. No INIT was withdrawn from Initia by the attacker.
We intend to restore the backing of affected assets in full. We are finalizing the plan and will share the details once it is confirmed.
As a precaution, we are rebuilding the affected infrastructure from scratch in an isolated environment, enforcing least-privilege access controls, and hardening security at every layer of the stack from cloud infrastructure to onchain IBC validation. Given the above, we expect this process to be longer than a routine network restart. We will share timelines as they become available.
All announcements will be made through Initia’s official channels. Thank you for your patience.
70
Coinbase gave AI agents a /feedback endpoint to report what's not working and what could be improved
and it seems the agents got what they wanted from their human servants in less than 24 hours!
pretty fun loop. an agent flags what's slowing it down, the team fixes it, and every other agent using the tool benefits from the improvement and can suggest what to fix next
58
don't hear much about Perplexity and its Computer, but it seems like they could be useful (for enterprise)
Persistent Agents are the next unlock of AI usage
Our spend on AI is now 45% persistent agents
Unfortunately for Anthropic, Claude Tags sucks
OpenAI's Slack agent is even worse
We primarily use Perplexity Computer, which uses a lot of different models
Perplexity's Harness mogs
8
388
Bitduke retweeted
Persistent Agents are the next unlock of AI usage
Our spend on AI is now 45% persistent agents
Unfortunately for Anthropic, Claude Tags sucks
OpenAI's Slack agent is even worse
We primarily use Perplexity Computer, which uses a lot of different models
Perplexity's Harness mogs – at San Francisco, CA
United wants your money, but doesn’t want your ai agent booking the flight
so United is restricting unauthorized third-party agents, even through travel agencies and aggregators, while other companies are opening their services to them
obviously, account security matters, but comparing fares, baggage fees and schedules is exactly what i want an assistant to handle, focused on my budget and the numbers without emotions. i’d rather see here secure integrations with spending limits and purchase approvals than only full restrictions
1
1
63
another sign Carbon is getting closer to those of us without Ultra, with more pro users reporting access through Google Search’s ai mode
The launch of Gemini 4 Flash, aka Hydrogen, appears to be imminent. If you don't have an Ultra subscription and are still waiting for a chance to try out Gemini 4 Carbon, I'll tell you a secret: Carbon is live right now in Google search AI Mode.
1
170
huh, United is fighting progress by blocking AI agents from booking flights, while companies like Shopify and Stripe are opening up shopping and payments to them through proper checkout integrations and APIs
idk what their reasoning is, but imo it's pretty clear who has more to lose from this
United Airlines is banning personal AI assistants from booking flights, arguing the bots pose privacy threats and may violate FAA rules. 🔗 on.wsj.com/489bKZJ
1
71
years of security verification compressed into four weeks, with more open proof work for other (ai) researchers to build on
@GoogleDeepMind Alphaproof code (an ai verification framework) proved specific security guarantees of the encryption code behind signal (the private messaging app). its agents generated over 6 million lines of lean proofs, with automated checks letting experts focus their review on roughly 1% of the codebase. in four weeks, they verified protections including keeping messages secure before a temporary device compromise and after recovery
also really important that deepmind is open-sourcing the verification artifacts, so researchers and ai agents can inspect, reproduce and build on the work
pretty exciting to me, beyond verifying signal’s security, useful parts could be reused to make future verification projects easier
For the first time, AI has mathematically proven @signalapp’s protocol is secure down to the code level.
Using our breakthrough AlphaProof Code framework, we verified the messaging app’s cryptographic system – which protects billions of private messages every day.
1
52
interesting paper from Unify, does doing more work actually make them better at the next job?
their benchmark brings back the same types of visual puzzles with fresh inputs, + new ones, to test whether agents remember what they learned and can apply it elsewhere
and the results are quite interesting, none of the tested memory-based setups showed measurable gains on unseen puzzles, while training Qwen3.5-9b to imitate a stronger model pushed its success rate from 17.8% to 62.5% in the four-user setup
so saving experience in notes and skill files DIDN'T automatically make these agents better at new problems, while updating the model itself did. that makes me curious about coding agents, could training on past bug fixes help them handle unfamiliar bugs with less guidance?
obv, this paper doesn’t answer the coding question, but i think it gives a useful way to test whether an agent actually gets better from experience, and makes me less convinced that adding memory alone is enough
The paper, the code, and replays of every run:
Paper: unify.ai/papers/continual-ar…
Code: github.com/unifyai/continual…
Site: continual-arc.unify.ai
47
a $1.4T valuation and investors are still lining up for Claude
Anthropic's IPO is already multiple times oversubscribed (per Bloomberg), with the IPO and a separate private placement together targeting up to $90.4 billion, so there's clearly strong demand even at the proposed valuation!
such strong demand at an enormous $1.4T valuation clearly shows how bullish investors are, and i wonder if some are already betting on another 2x from here, like what if Anthropic reaches $3T in a year?
1
54
Claude now leads 26% of model R&D research
so researchers can spend less time implementing experiments and more time testing ideas they previously couldn’t justify spending weeks on, pretty exciting
Sam McCandlish said five days ago that Anthropic would be releasing a report on the speed of internal progress, so we will finally get an update to this chart. I’m very interested to see what this number is now. METR thinks it is at 50%. AL4 is the last step before full RSI.
3
133
Bitduke retweeted
STATE OF THE TECH JOB MARKET IN 2026
Twice a year, I dig into trends in hiring and job openings. This is the fifth report I've done, and the most surprising yet:
1. PMs are winning in the AI era
2. Engineering demand has stalled
3. Design openings are at a record low
4. Every role now requires AI skills
5. Tech jobs are concentrating in the Bay Area
6. Remote roles are at a record low
7. The tech job market is growing fast, but mostly outside of software (gasp!)
Full report here: lennysnewsletter.com/p/state…
