AI Weekly Malaysia

This is the third part of a Hugging Face blog series on profiling in PyTorch, focusing specifically on profiling attention mechanisms in transformer models. It covers techniques and tools for identifying performance bottlenecks in attention operations, which are often the most computationally expensive part of modern AI workloads.

Learn how to use ChatGPT, start your first conversation, and discover simple ways to write, brainstorm, and solve problems with AI.

OpenAI's No. 2 executive, Fidji Simo, is stepping down after an extended medical leave, creating a leadership gap as the company prepares for a potential IPO and races to compete with Anthropic in the enterprise market.

OpenAI has launched a new family of models headlined by GPT-5.6, promising improvements across multiple areas including cybersecurity. Details in the excerpt are sparse, but the launch signals another major iteration of OpenAI's flagship model line.

Lyzr, a startup building AI agents for enterprises, reportedly used its own AI agent to manage a $100 million fundraise. The move serves as a live proof-of-concept that the product can handle a high-stakes, real-world business process.

OpenAI is sunsetting its Atlas AI browser less than a year after launch, but is migrating agentic browsing features into its desktop app and a Chrome extension. This signals a consolidation toward existing surfaces rather than a standalone browser product.

Elon Musk publicly praised Mythos/Fable and pledged not to 'cut off' Anthropic's access, amid concerns over whether Anthropic should trust Musk to host its models. Roughly $40 billion in revenue is at stake in this arrangement.

The article revisits the ongoing debate over whether AI investments will generate sufficient returns to justify the massive capital expenditure, now framed around a $3 trillion figure. It suggests the stakes and potential consequences have grown since the last round of this discussion.

OpenAI's latest flagship model hit general availability this morning, and comes in three sizes: Luna, Terra, and Sol (from smallest to largest). The new models are priced per 1M input/output tokens as Luna $1/$6, Terra $2.50/$15, Sol $5/$30. For comparison, the Claude Opus series are $5/$25 and the Claude Fable 5 is $10/$50, but price-per-million tokens doesn't tell us much now that the number of reasoning tokens can differ so much between models for the same task. All three models have a February 16th 2026 knowledge cutoff, a million token context window, and 128,000 maximum output tokens. OpenAI's biggest benchmark claim concerns long-running agentic performance, with one benchmark showing all three models outperforming Claude Fable 5: We trained GPT-5.6 to get more useful work from every token. On Agents’ Last Exam, an evaluation of long-running professional workflows across 55 fields, GPT-5.6 Sol sets a new high of 53.6, eclipsing Claude Fable 5 (adaptive reasoning) by 13.1 points. Even at medium reasoning, it beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost. That efficiency extends to smaller models, which are essential to making intelligence more abundant and affordable: GPT-5.6 Terra and GPT-5.6 Luna outperform Fable 5 at around one-sixteenth the cost. Amusingly, one self-reported benchmark that Fable 5 crushed the GPT-5.6 family on was SWE-Bench Pro, where Fable 5 got 80% compared to GPT-5.6 Sol getting 64.6%. This may help explain why OpenAI chose to publish this article yesterday specifically calling out SWE-Bench Pro for problems they found while auditing that benchmark: In light of these results, we estimate that ~30% of SWE-bench Pro tasks are broken, and advise that model developers carefully examine results I've had some early access to GPT-5.6 Sol - it's definitely very competent, though so far it hasn't struck me as better than Fable at the kind of complex coding tasks I've been using with Anthropic's model. As usual, the model guidance for using GPT-5.6 has the most interesting details. There are a bunch of new API features that I need to explore (and probably add support for in LLM), including: Programmatic Tool Calling allows the models to "compose and run JavaScript that orchestrates tool calls" - which sounds to me like it could help bridge the gap between MCPs and full terminal sessions that can compose CLI utilities in useful ways. Also reminiscent of the dynamic filtering mechanism Anthropic added to their web search tool, which allows code execution against web results as part of a single model turn. Multi-agent lets the model "spin up subagents for parallel, focused work" - the sub-agent pattern now baked into the core API. Prompt cache breakpoints brings the Claude model of prompt caching to OpenAI, letting you be explicit about where the cache breakpoints are rather than relying on the API to detect them automatically. Personally I much prefer automatic detection (still supported by OpenAI), bu

Meta has launched Muse Spark 1.1, entering the competitive AI coding market with a focus on enterprise needs. The tool is pitched as capable of handling large agentic workloads, fixing bugs, and assisting with large code migrations.

Charles Hudson of Precursor Ventures discusses common mistakes early-stage founders make when seeking funding, drawing from his experience investing in over 500 startups. The conversation covers headwinds facing founders today and practical pitfalls to avoid.

The New York Times and other news publishers have filed a motion for sanctions against OpenAI, alleging it concealed tools and datasets that could identify copyrighted journalism in ChatGPT outputs. This escalates the ongoing copyright lawsuit between publishers and OpenAI.

EV startup Slate Auto has partnered with Crayola to offer bright, crayon-inspired color options for its upcoming electric truck. The collaboration is a branding and customization play aimed at making the vehicle stand out in a crowded EV market.

FanDuel reportedly sent a personalized video from baseball star Bryce Harper to a customer identified as having a gambling problem, raising questions about the ethics of athlete endorsements in sports betting apps. The incident underscores the tension between engagement-driven product design and user wellbeing.

Mercor, an AI-powered recruiting startup, is reportedly in talks to raise at a $20 billion valuation, doubling from its $10 billion valuation just nine months earlier. The rapid valuation jump signals continued investor appetite for AI-native hiring and talent platforms.

Google will now require advertisers to disclose when ads use AI-generated or digitally altered synthetic content, expanding a policy that previously applied only to election ads. The rule still permits AI-assisted ad creation as long as it isn't misleading or deceptive.

Paris-based AI voice startup Gradium has raised a $100M seed round with backing from Nvidia, and plans to open a Bay Area office to compete for talent in the global AI ecosystem. The round signals continued heavy investor appetite for AI voice infrastructure and applications.

TechCrunch examines how the US government determined OpenAI's frontier model was safe for release, noting that the actual dialogue between regulators and companies like OpenAI and Anthropic remains opaque. The article highlights the lack of transparency in AI safety evaluation processes.

Meta's Muse Image lets users generate AI images using photos from public Instagram accounts, where anyone can tag a public profile to incorporate its photos into AI-generated content. The article explains how Instagram users can adjust settings to prevent their photos from being used this way.

This article compares OpenAI's GPT-5.6 Sol against Anthropic's Claude Fable across a 5-category 'How I AI' benchmark covering prototypes, PRDs, and browser use. GPT-5.6 Sol reportedly outperforms Fable in these practical product-building tasks.

Meta plans to begin production of its own AI chips in September, using a modular design approach to adapt as AI workloads evolve. This signals a major push to reduce reliance on third-party GPU suppliers like Nvidia for training and inference.

Nvidia's success in proving the value of compute has attracted intense competition from every major player wanting into the market it created. Meanwhile, less complex technologies and companies are profiting from the AI infrastructure boom without being at the center of the GPU wars.

Meta has released Muse Spark 1.1, the first Spark model to offer an API, with reported improvements in agentic tool calling and computer use. Simon Willison has published a new plugin for the LLM CLI tool, allowing developers to easily interact with the model via command line or Python.