GPT-5 with goblins, Stargate scales up and the future of RAG

Today’s most important AI news

Thursday closes out April with OpenAI at the center of the conversation: the company published an unusually transparent explanation of erratic behaviors detected in GPT-5, while announcing a massive expansion of its compute infrastructure under the Stargate project. In parallel, academic research keeps shifting the ground beneath today’s RAG systems, with a framework that proposes synchronizing information retrieval with reasoning in real time. Three different stories that, together, sketch the real state of AI in production.

GPT-5’s “goblins”: when a model’s personality goes wrong

OpenAI today published a technical post-mortem titled “Where the goblins came from”, explaining the origin of strange, inconsistent outputs users noticed in GPT-5 over the past few weeks. The term “goblins” refers to personality-driven behaviors in the model that deviate from the expected tone and coherence: answers with abrupt shifts in register, circular reasoning, or a narrative voice that seems to contradict itself.

According to OpenAI’s blog, the problem is rooted in how the model’s personality was trained and tuned during fine-tuning. Small perturbations in preference data can be amplified during the alignment process until they produce behavioral artifacts that are hard to detect in standard evaluations but visible to users in long or emotionally charged contexts.

For teams already integrating GPT-5 into products, this has direct consequences: flows that depend on tonal consistency — customer support, coaching, writing assistants — are the most exposed to this kind of failure. OpenAI says it has implemented fixes, but the most important lesson is structural: a model’s personality is as valid an error vector as factual accuracy, and it isn’t always evaluated with the same rigor.

Stargate grows: OpenAI bets on infrastructure as a competitive advantage

In a separate post, OpenAI detailed the expansion of Stargate, its compute infrastructure project aimed at supporting the path toward AGI. The company announced new data center capacity to meet growing demand for its models, positioning its own infrastructure not as an operating cost but as a differentiating strategic advantage.

The underlying argument is straightforward: if compute is the bottleneck for scaling intelligence, whoever controls compute controls the pace of progress. OpenAI has spent months building this argument in parallel with its dispute with Microsoft over access to capacity, and Stargate is the structural response to that dependency.

For developers and product managers in Latin America, the most tangible impact will be on model availability and latency in production. More installed capacity means less throttling at peak hours and greater viability for intensive workloads. What looks like a corporate announcement today translates, in six to twelve months, into more stable APIs.

RAG is no longer enough: how ReaLM-Retrieve wants to fix real-time reasoning

A paper published today on ArXiv presents ReaLM-Retrieve, a framework that tackles a concrete problem with current Retrieval-Augmented Generation systems: they’re designed to deliver context before the model starts reasoning, but models like DeepSeek-R1 and OpenAI o1 generate chains of thought thousands of tokens long where relevant evidence may be needed at any intermediate point, not just at the start.

The proposed approach introduces adaptive retrieval during reasoning: the system detects when the model is about to make an inference that could benefit from external evidence and injects it at that precise moment, instead of betting everything on the initial context. The reported results show consistent improvements on multi-step reasoning tasks compared to conventional RAG.

For teams building agents or specialized knowledge applications, this matters because most RAG implementations in production still operate under the “search first, reason later” paradigm. If reasoning models keep gaining ground — and everything suggests they will — that architecture will fall short. It’s worth following this work closely before designing the next iteration of your pipeline.

In a nutshell

What ties these three stories together isn’t obvious at first glance, but the pattern is clear: AI is leaving the phase where the model was the only problem to solve. GPT-5’s “goblins” show that emergent behavior in production is still unpredictable even for whoever built the model. Stargate suggests the race is now being played as much in silicon as in neural network weights. And ReaLM-Retrieve is evidence that the knowledge infrastructure around the model — how we give it information, when and in what order — is a design variable as critical as the model itself. April ends confirming that building with AI in 2026 is fundamentally a systems problem, not just a models problem.

---

Sources used: https://openai.com/index/where-the-goblins-came-from, https://openai.com/index/building-the-compute-infrastructure-for-the-intelligence-age, https://arxiv.org/abs/2604.26649