Martin Alderson

Finally got round to starting a blog, if you're interested in AI and software engineering I hope you enjoy it: martinalderson.com. And feel free to subscribe to my once a month max newsletter there too!

Why Claude's new 1M context length is a big deal martinalderson.com/posts/why-clau…

me (mar 7): "I think we'll see 'double usage limits' overnight" anthropic: "From March 13, 2026 through March 27, 2026, your five-hour usage is doubled during off-peak hours" good to see anthropic actioning my blog efficiently ;)

Did not expect this! Max/Business users can use 1M context on Opus 4.6 on Claude Code

Yes I get that, but why can't agents build those features? Also a lot of the complexity goes away when building for one org - eg only one auth system, certain workflow features, etc. I def can see orgs of the future having excellent purpose built eg product management software being a competitive advantage

A good part of building a product like JIRA is the nonfunctional and no viable stuff. The stuff devs using it don’t see but what makes it useful + sticky. Things like reporting, custom fields, data retention + SLA guarantees + disaster recovery, data governance, permissions etc etc etc

Cannot wait for teams that build their custom JIRA/Workday replacement/custom CRM to one day turn around and ask: “Why do we have so much internal software that is buggy / has poor UX + we need to maitain?” Seen this movie well before AI, when Uber built uChat (custom Slack)

@GergelyOrosz I'm curious though. How far would that few person dev team _with agents_ have got? Would it be subpar? I genuinely don't know the answer but it feels like you could get a lot lot further now than before. Esp for anything that needs a lot of "per customer" customisation?

The end of the Uber story: after 3 years of a few-person dev team building what was a subpar Slack: they moved to Slack. This is the thing with “rebuilding JIRA:” you don’t realize you forgot to build eg auditing, backups, disaster recovery etc until you actually needed them yesterday

Thought-provoking post from @martinald. One of my favourite writers on AI martinalderson.com/posts/is-the-a…

No, anthropic isn't losing $5,000/user on Claude Code subscriptions: martinalderson.com/posts/no-it-do…

Wrote some thoughts up on the compute crunch. Will be interesting to see uptime from Anthropic next week... martinalderson.com/posts/is-the-a…

@awnihannun Playing around I get qwen3.5 4b w/ 128k token context in just over 10GB w/ Q4 K quantisation, so you're right it would squeeze in there. But I'm not sure the inference speed would be usable at that length. It is genuinely amazing so much intelligence can fit in ~3GB tho!

@martinald I haven’t done the calculation (would be very curious to see it). But I would wager you could make >100k context length work on a phone with 10GB with KV cache quantization.

According to benchmarks Qwen3.5 4B is as good as GPT 4o. GPT 4o came out ~2 years ago (May 2024). Qwen 3.5 4B runs easily on modern mobile devices. So the gap between frontier intelligence in a datacenter and running a model of equal quality on your iPhone could be 2-3 years. (Probably closer to 3 assuming Qwen3.5 4B is more benchmaxxed than 4o) I don't expect the trend of increasing intelligence-per-watt to change. So in 2-3 years it's plausible we will be running GPT 5.x quality models on an iPhone. Pretty wild.

made a hook that adds a bouncing dvd logo to claude code whenever it's thinking

@trq212 Can we chillax with the WSL2 limitation? It should work fine (if you install pulseaudio, it sets itself up correctly). probably should just check if wsl2 has audio devices rather than just saying it's not supported?

Voice mode doesn’t cost extra to use, and tokens for voice transcription don’t count against your rate limits. Available on Pro, Max, Team, and Enterprise on a rolling basis.

Voice mode is rolling out now in Claude Code. It’s live for ~5% of users today, and will be ramping through the coming weeks. You'll see a note on the welcome screen once you have access. /voice to toggle it on!

Don't overlook how important the harness is too. The fastest token is one you don't need to infer at all. Think this is probably the most underdeveloped place to optimise. Eg Claude Code uses Haiku to parse bash output and detect if a file exists that it's looking for, so the agent can preload the file contents. This saves an entire turn on opus. Plus obviously all the prompt caching stuff. But I suspect there is so much more that can be done on the agent side to save tokens.

Inference compute is on track to be a massive computational workload by the end of this decade. I think it will be much bigger than training (especially if you consider RL rollouts / inference needs for training). And it's still an open playing field in terms of the hardware, the platforms, and the models. It's also increasingly clear that people are willing to pay a premium for reduced latency. On the hardware side there are several interesting directions to keep an eye on: - SRAM style setups seem promising (GPT Spark on Cerebras, Groq acquisition by Nvidia) - Disaggregated systems (prefill on one machine / processor, generation on a different one) probably make a lot of sense. The computational characteristics of prefill vs decode are so different, specializing at the hardware level will yield efficiency gains - I also wouldn't discount more exotic technology like the Taalas chip / near memory computing / etc. While they are still pretty far out from large scale deployment, the economic pressure for efficiency gains could be a catalyst On the algorithm / architecture side: - Pretty much every major open-weights model has at least one optimization which makes it faster for inference. Whether it be MoE, SSM (or other hybrid variety), or sliding window or sparse attention. There are more differences here than there were a year ago. And it will be interesting to see where we converge. - Will diffusion models unify the prefill / decode split? - Still believe there are big gains to be had in further co-design of model to hardware and workload I also don't think we will have a one-size fits all solution in the future: - Cloud-based models may look very different than edge-optimized models - Models may be more and more co-designed for the hardware they are deployed on - There will be at least one knob which trades-off latency and power efficiency / cost.

@nityeshaga No this has been around for a while. I think they removed it and re added it at some point

Just noticed something new in Claude Code's /compact command. Now in the compacted summary, the summary explicitly provides location to the full chat JSON file and tells Claude to reference it if it wants more detail. I've been looking at the compacted summaries since always, but this is the first time I'm seeing this prompt. I suspect that this is why compacting seems to be working more flawlessly recently within Claude Code.

Very interesting. Think we are going to see a lot more of this soon.

@DnuLkjkjh yep, makes sense and i suspect it is why apple is having so many problems with their "siri" replacement

spot on. the gap between 'model runs on device' and 'agent works reliably on device' is enormous. tool calling, memory management, error recovery — all of that needs to work without a round trip to a server. I've been building on-device voice processing and even for a narrow domain like transcription, the edge cases multiply fast when you can't fall back to cloud

Wrote some thoughts on why on device agentic AI is proving so elusive: martinalderson.com/posts/why-on-d…

The entire Accenture workforce is about to be outperformed by a 24-year-old who learned Claude Code last Tuesday.

Some really insightful thinking from @rowlsmanthorpe on the potential labour market disruption rowlandmanthorpe.substack.com/p/the-case-of-…

Found a cool pattern for running AI code reviews with @opencode, without a seperate SaaS subscription. Details here: martinalderson.com/posts/using-op…