GPT-4o sees but doesn't understand: AI's vision problem
Multimodal models fail at basic computer vision tasks, OpenAI moves into finance and RAG needs less noise.
Today’s most important AI news
The day was marked by research that casts doubt on capabilities many take for granted: the most popular multimodal models — GPT-4o, Gemini, Claude — have serious gaps when asked to do real computer vision, not just answer questions about images. Add to that an alliance between OpenAI and PwC aimed squarely at the finance function of large companies, and a paper that rethinks how RAG should be designed when the reader isn’t a human but a model.
GPT-4o can describe a photo, but it can’t segment or detect like a specialized model would
A benchmark published on arXiv evaluated the main multimodal models — GPT-4o, o4-mini, Gemini 1.5 Pro, Gemini 2.0 Flash, Claude 3.5 Sonnet, Qwen2-VL and Llama 3.2 — on standard computer vision tasks: semantic segmentation, object detection, image classification, depth estimation and surface normal computation. The results are uncomfortable: performance on these tasks falls far short of what smaller specialized models achieve effortlessly.
The underlying problem is conceptual. These models were trained to answer natural-language questions about images — and they’re remarkably good at that — but the visual understanding involved in segmenting every pixel of a scene or precisely estimating an object’s geometry requires something different. It’s not that the models “see badly”: it’s that the kind of understanding they developed isn’t the same as what classic computer vision demands.
For developers evaluating whether to use GPT-4o or a similar model as a replacement for a computer vision pipeline, this research is a concrete warning sign. If your use case involves object detection in production, medical image segmentation or geometric analysis, today’s frontier models aren’t the right tool yet. Combining a specialized model for perception and an LLM for reasoning remains the most robust architecture in these contexts.
OpenAI and PwC go after the finance function of large companies
OpenAI announced a collaboration with PwC to bring AI agents to the CFO function at large companies. The stated goal is to automate financial workflows, improve forecast quality, strengthen internal controls and, more broadly, “modernize the office of the CFO”. PwC would act as implementer and advisor, while OpenAI provides the model infrastructure.
The move is significant for what it signals about the direction of the enterprise market. Until recently, AI use cases in corporate finance were limited to report automation or structured data analysis. Entering the CFO’s territory — where planning, control and compliance decisions are made — is a leap in complexity and regulatory risk. PwC gives OpenAI something no model can generate on its own: institutional credibility with boards and auditors.
For product managers and executives at mid-size and large companies in Latin America, this has practical short-term implications. The big consultancies are already packaging these solutions, which means the conversation about AI-driven financial automation will stop being optional in many organizations. Understanding what these agents can and can’t do — especially in local regulatory contexts that differ from the US market — will be a real competitive advantage.
RAG needs a redesign: when the reader is a model, noise is more dangerous
A perspective paper published on arXiv makes an argument worth taking seriously: information retrieval systems were designed with human users in mind, who tolerate a certain level of noise and can ignore irrelevant results. When the consumer of that information is an LLM, the rules change completely. Models have limited attention budgets and are especially vulnerable to noise: irrelevant or misleading information isn’t a minor inconvenience, it’s a direct cause of hallucinations and reasoning errors.
The paper’s central proposal is to adopt a “denoising first” perspective: before passing context to a model, the retrieval pipeline should actively filter out noise, not just rank by semantic similarity. This means rethinking metrics, reranking architectures and the way search indexes are built when the end consumer is an agent or an LLM in a RAG flow.
For developers building RAG systems today, the takeaway is immediate: context quality matters more than quantity. A retriever that returns ten mediocre fragments is more harmful than one that returns three precise ones. Optimizing the retrieval pipeline with explicit filtering logic — not just embeddings and cosine similarity — is probably the highest-impact improvement you can make to an existing RAG system without touching the base model.
In a nutshell
Today’s pattern is a tension between expectation and reality. Multimodal models promise complete understanding of the visual world, but fail at that domain’s most structured tasks. RAG promises enriched context, but without active denoising it introduces more problems than it solves. And OpenAI’s entry into CFO territory promises financial automation, but it leans on a traditional consultancy to gain institutional legitimacy. The AI of 2026 remains enormously capable and, at the same time, fragile at the edges. The professionals who understand exactly where those edges are will be the ones making the best adoption decisions.
Sources used: https://arxiv.org/abs/2507.01955, https://openai.com/index/openai-pwc-finance-collaboration, https://arxiv.org/abs/2605.00505