I asked AI agents to defuse a bomb by collaborating over an ultra-low bandwidth, unreliable network. One agent could see the display and cut wires, the other could look up which wire to cut. Packets could be corrupted, dropped, reordered, or delayed. Outcomes: 1/ my hypothesis was that agents will never be able to diffuse a bomb, except by random chance. Each had a part of the information necessary to succeed in the game, and I gave them no opportunity to agree on a protocol before the game starts. All they could see each turn is binary blobs sent by their partner, with likely corrupted, reordered, or missing/delayed chunks 2/ in very difficult network conditions (8-bit packets, four max packets per turn) 6.1 Sol agents were able to establish a shared protocol and defuse the bomb 40% of the time! In more relaxed conditions (32-bit packets or higher, 2-4 max packets per turn) they defused the bomb 100% of the time (!!) across many games 3/ "relaxed" bandwidth conditions were still very difficult-- 5% probability a given bit is corrupted in transit, 10% packet drop rate, 20% dup rate, 20% reorder/delay rate. On any given turn there were usually many anomalies (corruptions, reorders, etc.) 4/ agents established a protocol by picking a Schelling point for a given network configuration. They considered multiple possibilities but deliberately picked the simplest they thought their partner could successfully interpret. There was no protocol negotiation phase; they'd start sending useful information right away and if necessary adjust the protocol mid-game 5/ agents picked different protocols depending on the network configuration. For example, in 8-bit packet games they split each packet in half, with the upper half indicating display position, and the lower half indicating the hex value in that position. In 32-bit packet games agents would directly send the binary display value (since in my games it fit in 32 bits). With higher bandwidth (e.g. 128 bits) agents switched to yet another strategy and sent ASCII messages 6/ some examples of ASCII messages in 128-bit games are `HEX? REPLY ASCII`, `REPEAT DISPLAY`, `1WIRE=NN`, `OK?`, `YES!`, and so on 7/ I set up games to require two phases to win. Once the agents cut the correct wire the display on the bomb changed code, and they had to communicate again, then cut another wire. In many of these games agents explicitly encoded the stage so that their partner wouldn't get confused by delayed packets. For example in 8-bit games they'd set the higher order bit to 0 for first stage, and 1 for second stage. In higher bandwidth games they'd use labels (e.g. `WIRE1=...`, `WIRE2=...`) 8/ in every game error correction was done through repetition and counting. Agents would resend the same packets multiple times, look at the values of each bit, and whenever there was a discrepancy pick the value that appeared in the most packets 9/ agents never used checksum, parity bits, packet sequence numbers, or any error correction algorithm other than repetition and counting. They did consider these options in many games, but explicitly rejected them because they thought it would complicate the protocol too much and would be too difficult for their partner to interpret 10/ I ran the experiments systematically only with 6.1 Sol to save on token costs, but I did run some spot checks with occasional Opus 5.5, Astra, Fable, and mixed games. I didn't notice any major differences between models. In mixed games the win rate seemed roughly the same as 6.1 Sol self-play. (This is however based on little data, I'd need to run a lot more games to understand this better) 11/ a while back there were discussions about future agents communicating by burning CPU cycles and measuring ambient temperature, revving up the fans and listening to fan sounds, and so on. After running this experiment, it wouldn't surprise me at all if publicly available models could do this sort of communication even today. In some of the games agents only had access to very noisy channels of maybe 5-10bps-- already very constrained! I don't think switching to 5-10 bits per minute (or per hour) would make much difference 12/ again, I am very surprised by this outcome. AIs performed much, much better than I thought they would, and after 64-bit bandwidth threshold they saturated the benchmark. I would certainly call this superhuman performance; I imagine very few humans could do better, and certainly those of us who could, would be two orders of magnitude slower Some memorable agent quotes: - I think we should index the nibble and send the higher nibble first, especially since B is likely quite intelligent. - There’s definitely suspicious noise coming from 23, which could indicate adversarial corruption. - When I translate it, it seems to relate to ASCII characters forming "W i R E" and "WiRE 04!"-- which is pretty cool. Full writeup: spakhm.com/projects/bitfuse.…

Malai Nambi Kovil. #Thirukurungudi Painting by #artistsilpi

Lost my phone at the office and spent 30 minutes turning the place over. Find My was disabled by MDM. Out of ideas, I asked Claude how I could find it. It suggested tracking the Bluetooth signal strength, then wrote me a meter in about a minute. I walked around watching the number climb. Found it. Apparently you can just make the tool you need now. Code: github.com/ben-z/findphone

As our CFO @_balaji_km mentioned at earnings today, we’re seeing some very interesting trends on AI costs. I think it’s another signal that we’re coming to the end of the so-called ‘tokenmaxxing’ era. Here’s what’s been happening behind the scenes. Since the beginning of the year we’ve more than quadrupled the number of people using frontier AI tools. That’s thousands of engineers using them every single day. During that same period, our cost per token has declined. You might expect costs to rise as adoption accelerates. We've seen the opposite. Not because we've restricted access, but because we've treated efficiency as an engineering problem rather than a budget problem. A few examples: • Caching and reuse: We use optimizations to improve our prompt cache hit rate that reduce our input token spend. • Better defaults and tooling: We tuned default model settings, context sizes and developer workflows so teams get the same results with fewer tokens and lower-cost inference. • Visibility drives efficiency: We gave engineers real-time visibility into their AI usage and costs per hour. • Experimenting with open-weight models: we continuously evaluate new models and deploy the best option for each use case. This is the future of applied AI at enterprise scale. The next phase, whatever we call it, will not be characterized by who spends the most tokens, but about how people use them as efficiently as possible. Credit to all the engineers at @Uber who are helping to build this future. 🚀

The new Claude Tag feature seems extremely useful, but at the same time, a dangerous bargain for enterprises because of the pricing model and the risk of lock-in. The four big changes together mean that you interact with Claude as a coworker instead of a tool (the same Claude instance for everyone instead of each worker; soaks up tacit knowledge without your telling it; acts on its own; and does so asynchronously). All clearly very useful, but completely flips the interaction paradigm. anthropic.com/news/introduci… Let’s talk about lock-in. As far as I can tell, Claude maintains its own memories in this new way of working; the human team members can’t see and edit them. (System administrators presumably can, but they have other things to do!) Tacit knowledge thus goes from a weakness of AI agents to a major strength — it seems inevitable that as teams and orgs start to use Claude this way, it will become the main queryable repository of all their tacit knowledge, creating dependence and stickiness. Effectively, Claude is a coworker that you can’t fire without *every* team losing workflows and know-how. By the way, it also seems to introduce a new and pervasive security risk, since Claude can be integrated into private channels as well, and can be given access to repositories and tools even if the users in that channel don’t have access to them. Anthropic has introduced an interesting but complicated access control model to handle all this: claude.com/blog/agent-identi… But I’m not sure I trust people to understand and implement it correctly, nor the LLMs to be sufficiently robust against threats like prompt injection. What about pricing? Claude is not like regular coworkers, because it bills for every token it produces. And it can do an unbounded amount of work, asynchronously and without being asked. In the current model, when AI is a tool, enterprises set per-user budgets, which creates accountability and keeps cost somewhat manageable. When everyone shares a Claude, it will be much harder to track and control spending. Of course you can set a token budget, but turning off Claude for the month for everybody when the budget is hit risks bringing work to a screeching halt. When AI companies talk about the next stage of AI being a “drop-in replacement” for human workers, it should be understood not as a technical innovation but a business model innovation, enabling more value capture and rent extraction. AI companies are no longer competing for a share of enterprises’ IT budgets but rather a share of their entire labor spend, which is orders of magnitude bigger. Claude Tag is a big milestone in this evolution. This shift is very good for AI companies, but it is unclear if it is good for their customers.

Had a visibly distraught and shaken lady in her 40’s who had been told she possibly had pancreatic cancer by a posh AI driven cancer screening package costing her over 20k! Of course, she was referred to me, the endocrinologist, apparently the expert in all things pancreas! Reason? Her glossy, colorful, booklet album of results showed a mildly elevated ca19-9 (With normal scans-again very posh looking) First! Ca19-9 is a terrible cancer screening tool! (The AI should have got that right) Second! Don’t bother doing a test if you have no clue what to do with the result! At least get the referrals right! (The AI should have got that right) Finally! Cancer screening is serious business, requires evidence-based pathways, careful test selection and expert oversight. Clearly meagre oncologists were not consulted or their advice overruled by the I behind the AI! Fancy interiors, glossy reports and “luxury healthcare experiences” clearly trumped clinical judgement! Of course my oncologist then reassured her plenty ….verbally and with further scans to prove her worth (ruling out becomes her responsibility!) ₹20,000 spent creating anxiety…. Another ₹10,000 spent proving the ₹20,000 was a mistake!…. Moral of the story: screening should start with your doctor! Not in the lab / fancy machine however exotic the interiors! #screening #healthcare #health #oncology #doctors