🔗 Learn more about me, my work, and how to connect: maeste.it – personal bio, projects, and social links.
A tough week for Anthropic, which had to admit to three major issues that sparked varied reactions in the community. Meanwhile, OpenAI gains ground on them with Codex and GPT-5.5, and on Google with ChatGPT Images 2.0, which competes closely with or perhaps surpasses NanoBanana. The Chinese are no less, with the arrival of DeepSeek version 4 and Qwen 3.6, all open source. And I wonder if it isn’t time to focus on open source tools instead, but read all about it in my deep dive.
Podcast with Alessio and Paolo:
In the latest episode we talk extensively about what I discuss in this week’s deep dive and show image examples
A great interview with Luigi Congedo is out, bringing his experience from American VCs and telling us about his courageous choices in Italy. Don’t miss it.
We’re trying to improve the audio/video format and podcast presentation... any feedback is welcome
Next Wednesday Stefano Gatti returns for an interview. He’ll share his current reading of hybrid intelligence that he anticipated in October, and this week of reckoning is a good context to revisit his perspective.
By now you know about our GitHub repository with tools and configurations for AI coding from the terminal on Linux. It now has its own site with single-script installation at Lince.sh
We released AntiVocale (Google Play, GitHub), a software to translate voice messages into text
On my own:
The video of the talk Alessio and I gave at VoxxedDay Zurich has been published
On May 30 I’ll have the honor of being one of the PyCon Italia speakers
On June 12 I’ll be in Catania as a speaker at Coderful
More talks at other conferences coming up...
This week’s news was dominated by two things for sure. The arrival of GPT-5.5 and ChatGPT Images 2.0, which marked a major step forward for OpenAI in both agentic AI and coding as well as image generation. The results are truly remarkable for OpenAI, because they’ve closed the gaps that until recently were quite evident with Anthropic on the agentic side and with Google on image creation. GPT-5.5 achieves state-of-the-art results on Terminal-Bench 2.0 (82.7%) and SWE-Bench Pro (58.6%), and for the first time an OpenAI model seems truly capable of competing with Claude Code in agentic coding.
But in parallel, the Anthropic case exploded. They admitted to having three significant bugs in Claude Code that affected the performance and quality of their flagship model Opus. The postmortem is detailed: on March 4 the default reasoning effort was changed from high to medium without announcement, on March 26 a cache management bug wiped the thinking context at every turn, and on April 16 a system prompt to reduce verbosity compromised the quality of generated code. All three issues were fixed by April 20, and Anthropic is resetting usage limits for all subscribers as a goodwill gesture. This confirmed a feeling already present in the community, which sparked various reactions, some perhaps excessive, calling the situation unacceptable and gravely serious. I honestly think that bugs in the software world can happen, and this won’t be the first or the last. On the other hand, I understand the position of those who complained that instead of having so many releases and new features, perhaps Anthropic should focus on the stability of their system, because it’s becoming increasingly central for a large mass of developers.
But if we look back at OpenAI’s house, maintaining an exaggerated release velocity seems to be a necessity, because otherwise it takes a moment to fall behind or get overtaken by another company. The good news, if we want to see it that way, is that some developers who decided to switch to other tools, Codex in particular, to try GPT-5.5, have certified how the use of standards, and in particular skills, allows an easy transition to different tools.
In this sense I’m glad I anticipated the use of multiple coding agents in LINCE.sh, because as we can see in this case too, tying yourself inseparably to one coding agent and one vendor might not be a good idea. And extending this concept a bit, I think the time has come for me to start looking at some open source agents. If state of the art models aren’t feasible locally or open source, or at least not yet, what we can do as AI engineers is to at least use tools that are completely under our control, and why not, maybe contribute to their improvement. By the way, it’s been a rich week for open models: DeepSeek V4 with 1.6T parameters and 1M token context in open source, and Qwen 3.6-27B, a dense 27B parameter model that beats its 397B MoE predecessor on all major coding benchmarks, show that the distance between proprietary and open source models is narrowing at least on architectures.
There’s an embarrassment of choice out there among open source tools, including for example Goose or OpenCode, but I think this week I’ll try to focus on PI and Hermes Agent. Two very different tools, one for coding and the other for generic agents, but both with very interesting features. The first is truly minimal and grows through extensions, and has been used very effectively to implement Karpathy’s autoresearch. The second is interesting because, doing very similar or identical things to OpenClaw, it has a truly remarkable attention to security and isolation of code generated and executed by the agent.
Official postmortem on three separate bugs that degraded Claude Code quality between March and April 2026.
A postmortem worth reading. It’s definitely interesting to see how Anthropic lays out the three errors, but it’s also very interesting to read how difficult it was to investigate the cause of these errors. The complexity is truly high in these systems, and not only because of the models.
GPT-5.5 brings state-of-the-art results on Terminal-Bench 2.0 and SWE-Bench Pro, ChatGPT Images 2.0 generates images with high-quality text, and Mollick tests the new capabilities in depth.
I mentioned in the deep dive how significant the advances at OpenAI have been. Here you’ll find the announcements of the new GPT-5.5 model and the connected tool Codex, which are doing truly important things, both from what I’ve read and from my own direct testing. And there’s also the announcement about images generated by ChatGPT Images 2.0. If you’re curious, go check out my podcast thumbnails, which are generated or retouched by ChatGPT Images. Don’t miss professor Ethan Mollick’s detailed analysis: absolutely a must-read article, and it’s also fun to check out his simulations to get a sense of how models have evolved in recent months.
DeepSeek V4 in open source with 1.6T parameters and 1M token context, and Qwen 3.6-27B, a dense 27B model beating its 397B predecessor on all coding benchmarks.
The Chinese are certainly not standing still, and the announcement of DeepSeek V4 in preview arrives. It’s an important announcement because DeepSeek, you may remember, about a year ago generated a real earthquake by demonstrating that open weight models could compete with American state of the art models. More or less the same thing is happening again, but what always makes DeepSeek releases interesting are the papers they release alongside them, which in the case of the previous version led to concrete advances not only for DeepSeek but for the entire community. This time’s paper seems very interesting too, though perhaps less disruptive, but I haven’t had the chance to dig into it enough for a detailed comment yet. Maybe I’ll come back to it. Meanwhile, Alibaba also releases Qwen 3.6-27B in fully open mode, and it’s a significant release because the benchmarks are truly remarkable and much better than even Gemma 4, which made so much headlines just a few weeks ago.
Enterprise platform to build, scale, and govern business agents, with low-code Agent Studio, Agent Development Kit, and over 200 models in the Model Garden.
You may remember that a couple of weeks ago Anthropic announced its version of an agent platform. Well, Google has already arrived with its own announcement too, and this identifies a fairly strong trend among big tech companies that are starting to provide proper platforms that greatly simplify the development of cloud-based agents. We may be seeing the emergence of a new trend that shifts agents, which today run locally on our machines, toward the cloud. I’m not sure, because part of the appeal of agents running on our machines is having access to our data, our systems. But certainly as we move into the enterprise world, we’ll see agent platforms like Google’s and Anthropic’s establishing themselves. I don’t know if we’re ready for the enterprise world yet.
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.