Claude — the new AI kid on the block

Curiosity, mostly. That's what got me to try Claude.

I'd been on ChatGPT long enough that it had become background infrastructure - always open in a tab, always useful enough to not question. But I kept seeing Claude mentioned in circles I pay attention to, and Anthropic's approach to AI safety struck me as more considered than most. So last month I started routing some of my everyday prompts through Claude instead, just to see.

A few things became clear quickly. The writing it produces doesn't read like a language model wrote it. (Well, it is.) That sounds obvious, but spend enough time with ChatGPT and you start recognising the cadence; a certain rhythm, a tendency toward lists when prose would do better, hedges stacked on hedges. Claude's outputs have more texture to them. Personality may be too much of a stretch. 

The other thing, and this surprised me more, is that Claude seems to know what it doesn't know. (Is this possible for an LLM?) ChatGPT will occasionally confabulate with remarkable self-assurance - detailed, plausible, and wrong. Claude is more likely to flag uncertainty, or to ask a clarifying question before charging ahead with an assumption. In a work context, that matters. Confidently wrong is often worse than uncertain. Of course, an LLM is an LLM is an LLM.

I should be fair: ChatGPT still has its edge cases where it pulls ahead. The plugin ecosystem is broader, and GPT-4's code interpreter has saved me hours on data wrangling tasks I don't particularly enjoy doing manually. These aren't minor points. And for the casual user, the difference probably won't be dramatic enough to bother switching.

But my default has shifted. When I open a new tab now, it's more often Claude than not. After months of the same habit, that's telling. 

(Title is a reference to an older post)

Breaking the Network Effect

 As most observers of Internet culture know, the power of a network is derived from the number of users on the platform. A single telephone user would have no utility from the device, but as more users join the network, the value to any single user goes up exponentially. This is also known as Metcalfe's law. 

All social media networks also follow the same principle - Facebook, Twitter, LinkedIn, and also WhatsApp. Most of these platforms have millions, if not billions, of users. As such, it's difficult for any new platform to challenge the incumbents. Even Google tried and failed multiple times to beat Facebook.

With all the recent news about updates to WhatsApp's terms of service, I had seen many friends moving to Signal recently. I personally installed Signal sometime back, but couldn't continue using it for the very same reason - network effect. If none of my friends are there, it isn't useful for me either. I uninstalled it soon after. 

However, a surge in migration provides opportunities to break the network effect that keeps the incumbent the leader. As a few people move to the new platform, they can choose to request all of their contacts to move as well. Many may not move the first time, but as they see more and more of their contacts moving and getting frequent requests to switch, the pressure builds until they also see enough value to move.

What can stop this from happening are switching costs - for example, WhatsApp has already opened up its API to allow businesses to interact with their customers.  As long as these businesses provide enough value for the WhatsApp users and don't replicate this on Signal as well, it will be difficult for the user to move and continue getting the same service. 

Personally, I have tried to convince some of my contacts to switch. Too early to say whether we will break the Network Effect.

Making WFH work

Redundancy. 

Ever since the COVID-19 pandemic hit the world and the entire family had to rely on the home Internet connection for getting everything done, I was dreading the day when that connectivity would go out. That day came a couple of months back, when my wife and I had work calls lined up and both the boys were about to start their online classes. Thankfully, that exclamation point near the Wifi icon disappeared in five minutes, but scared me enough to make me sign up for a second, back-up connection.

Some basic research led me to the venerable EdgeRouter X, which could take in both ISP connections and provide a single load-balanced connection to the wireless AP. Setup was simple enough, the folks at Ubiquiti have done an amazing job with the wizard that can get you started in less than 5 minutes. Basic networking knowledge is helpful too, to get more done out of this workhorse of a machine. 

What pushed me to the EdgeRouter was its Gigabit connectivity, whereas the cheaper TP-Link ones cap out at 100 Mbps. With one 150 Mbps and another 100 Mbps connection, I didn't want the router to be the bottleneck. Speedtests at fast.com and speedtest.net both provide around 230-240 Mbps, meaning that load-balancing is working, and is actually aggregating for multi-part connections. Also setup Tasker and Connectbot on my Android to SSH into the router and toggle load-balancing and failover at the touch of a button. 

Unless something drastic takes out both the connections simultaneously, it's a reassuring feeling to know that there's some redundancy to the lifeline that's making things manageable in these crazy times. 

Wireguard - the new VPN kid on the block

Having tried my hand with OpenVPN to provide a secure Internet connection for my gadgets while travelling, I was less than impressed by the throughput on the Raspberry Pi. At a measly 3 Mbps, it was simply not good enough.

I had heard about Wireguard almost a year back, but last week's ArsTechnica article rekindled my interest. Many others have also tried their hands, and looks like it's all set to take over the VPN landscape. Major benefits include a much simplified setup (compared to the maze that is OpenVPN), superior performance (throughput much greater than others), much smaller codebase (only 4,000 lines of code, compared to 600,000 for OpenVPN and 400,000 for IPSEC) which theoretically translates into a reduced attack surface and others.

Spun up a VPS instance and installed Wireguard. From install to setup on both server and client, it literally took less than 10 minutes and I had a working connection on my Android phone. A quick test indicates almost 4-5x times faster speeds than OpenVPN, and nearly instantaneous connection against the roughly 8-10 seconds for OpenVPN. Another major benefit seems to be that the protocol is not chatty, and packets need not be sent regularly to keep the connection alive - a major concern for mobile devices with batteries that drain quickly. It also promises full IP roaming on both peers (term to refer to both 'server' and 'client' on Wireguard), again significant when switching between WiFi and cellular connections where each transition would have otherwise initiated a full renegotiation of the session.

Still under heavy development, I couldn't find a package for the Raspberry Pi yet, although it can be compiled. The Android release is very basic and experimental. However, Wireguard could be soon integrated into the Linux kernel itself, giving it much needed code review and mainstream acceptance. Even Linus himself seems to be a fan.

Six second ads and consumers

Seems like the whole of advertising industry is moving towards shorter ad formats. Driven primarily by behemoths YouTube and Facebook, and taken up by other digital publishers, even TV is embracing six seconders as the next big thing. Is this just us admitting to the fact that consumer attention spans are rapidly shrinking or just that we can't be bothered to properly craft the story with the time and attention it deserves or maybe a combination of both?

There's definitely a strong case which can be made for the six-seconder. A brand which has already reached threshold levels in awareness may benefit by using it to amp up the frequency, and thereby, top-of-mind recall - major FMCG brands being an obvious example. There are also many brands with creatives which are worth watching over and over, in infinite loops of six seconds each. But when everyone jumps on the bandwagon, I'm not too hopeful the thinking will last, and we will end up being pummelled by them.

In a hyper-fragmented landscape of infinitely scrolling timelines and unlimited media choices, the primary benchmark by which creative and media efficacy should be measured is very simple - relevance. Are we reaching the right audience in the right context and telling her a story that she finds useful, and maybe even enjoys? Chances are, she will appreciate the effort and remember us for it. Or would we rather cop out and hope that she forgives us the six second intrusion, since there's so much less of it to hate?