Parameter Update: 2026-04

Bit of a slower week, but rumors are flying around OpenAI and Google having incoming launches soon, so I'm still hyped. Also: damn, the qwen team is shipping like crazy!

Qwen3-TTS

At least on my timeline, things have been pretty quiet around TTS models for a while. There were some interesting demos of voice cloning (if you haven't warned your parents about these - you probably should!), but for the most part, it seemed that ElevenLabs would hold the top spot for a long while. This week, that changed somewhat, with Alibaba launching a series of models around TTS generation that come dangerously close in quality - while being 100% free and open source. In my experience, audio quality is a bit lower, but the generated voice actually sounds better (take a listen below!). We even got a nice tech report paper. Nice!

ElevenLabs

0:00

/22.5175

1×

Qwen3-TTS

0:00

/17.497958

1×

What surprised me even more was the voice cloning capabilities of the base model. Below, you can listen to a reference audio I recorded using my Airpod's mic and an example generation:

Reference audio (real voice)

0:00

/6.314667

1×

Qwen3-TTS Voice Clone / Base model generation

0:00

/3.51175

1×

As is common with the Qwen team, they are also iterating rapidly on tooling and collaborations around their models:

Anthropic

Claude in Excel

After launching in October last year, Claude in Excel is now available on for all paid users, starting with the Pro plan. It is also more capable than before, supporting multiple input files, auto-compacting long conversations and overwriting existing cells less than before. The most astounding thing about this announcement is that, as far as I can tell, Copilot has still not caught up?

Claude Code in VSCode

Similarly, the VS Code extension for Claude Code has moved into general availability this week. When I last tested it, it still felt a big buggy, but that might be fixed by now? Either way, it must suck to be Cursor right now, with first-party CLIs eating your lunch and, if this is any indication, not leaving any crubs behind.

Constitution

In the past, Anthropic has stood out to me as being particularly concerned about alignment and model wellfare (including some recent blog posts), especially compared to the heavy monetization push OpenAI is going through right now. This week saw another installment of that, with them publishing a "Constitution document" - a document that serves as both a description of Claude's personality, value system and intended behavior, and a direct input in model training to ensure Claude will actually act as intended in the doc.

The whole thing is long, but well worth a read, and I am very interest in seeing if other AI labs have similar documents (and, assuming Anthropic stick with this one, how it might change in the future/affect future models). Either way, it's good to see this type of transparency!

Ralph Wiggum?

Not a purely Anthropic-focused news item, but given it's mostly centered around Claude Code, and Anthropic has since released official tooling for it, so I decided to include it here. For the past few weeks, my timeline has been filled with Ralph Wiggum. The core idea of Ralph loops: Run Claude Code in a while loop until it has completed the task at hand. Each iteration starts with a new context window - the codebase and self-written documentation are the only source of truth. Ideally, each loop leads to some form of progress (i.e., either moving the codebase forward, or failing at something leading to new insights for future iterations). In contrast to base claude, this can allow the model to complete much more complicated tasks (if your prompt is well specified) - on the other hand, it also burns an order of magnitude more tokens and is far from a magic bullet. Also, unfortunately, it seems the original "creator" has since pivoted to shilling crypto, so make of that what you will.

Clawdbot

The same list of people that were extremely excited about Ralph for the past few weeks have now picked up a new toy: Clawdbot.

You can think of Clawdbot as an open source, self-hosted (though still using Anthropic's models) Claude Code/Siri crossover that integrates with a bunch of messaging services, has memory, browser use capabilities and can be freely extended via MCP. If you're familiar, it's been compared to Poke:

On my timeline, people are buying Mac Mini's for the sole purpose of running this thing, and I've seen some extremely cool examples for the agent doing things:

That being said, for Clawd to work well, you're going to want to hand over a lot of permissions to it - some of which might leave you open for prompt injections or other types of abuse, so be careful:

Google

Personal Intelligence in AI Mode

Last week, I complained about Google stumbling in their launch of the "personal intelligence" feature in Gemini that better integrates information from other Google services into the assistant. This week, Google seemed determined to make fun of me specifically by announcing they would roll the feature out in AI Mode in search (meaning it will eventually reach over a Billion people). Nevertheless, the whole thing remains opt-in for now, so joke's on them.

OpenAI

Age prediction in ChatGPT

OpenAI has indicated in the past that they would like to experiment with more adult-oriented content. In light of recent age restriction laws sweeping across (at least) the EU and the US, this desire was always going to require some resolution. This week we saw how they plan to differentiate users: Have AI classify them, based on their ChatGPT usage behavior. For now, this is not rolling out in the EU, and I personally think it feels very invasive (plus, it's bound to be wrong a significant amount of the time), but if you're going to force age verification on people, this is at least slightly better than forcing everyone to upload their ID to weird online portals (looking at you, UK!).

Stargate Community plans

While I am still skeptical of the huge data center buildout plans OpenAI has announced with Project Stargate, this week, they gave us some insights into their plans to achieve them without destroying local energy grids. So that's good, I guess? Would really appreciate similar plans around RAM prices, right about now though.

Codex Agent Loop insights

Finally, last week brought us a very interesting blog post by OpenAI with a bunch of details on the Codex CLI agent loop. Well worth the read if you missed it so far. On the other hand, the whole thing is open source anyway, so if you're actually interested in learning more, go through the codebase with Claude (or Codex, if you want to be meta about it).