The bottleneck in software is no longer engineering hours

Beyond Coding · 2026-08-22

This video, featuring Michael Giannangeli, Head of Product for Agentic AI at Amazon Nova, discusses the shifting bottleneck in AI development. The challenge is no longer primarily about engineering hours, but about the high costs associated with training and deploying newer, more capable models. Giannangeli suggests using a simulated environment where AI models can rapidly iterate through trials and errors as a potential solution to mitigate these costs and accelerate development, especially for agentic AI.

More than just code review

Simon Willison · 2026-08-22 · 1 min read

Willison argues that effectively using AI coding agents requires two distinct skills: clearly instructing them on what to change, and confidently verifying the result — but that verification doesn't have to mean reading every line of generated code. This matters to senior engineers because it reframes code review instincts built around human-written code, suggesting you need to develop new validation strategies (tests, behavioral checks, diffs, etc.) rather than defaulting to exhaustive line-by-line review when working with agentic tools.

Which AI Models Are Worth Using

Theo - t3.gg · 2026-08-22

This video presents a tier list of current AI models, ranking them based on their utility, cost-effectiveness, and capabilities for software development tasks. The creator emphasizes the importance of web access for AI agents and highlights how models like Fable 5 and GPT-5.6 Sol excel in complex coding, while cheaper alternatives like GPT-5.6 Luna are ideal for auxiliary tasks. The tier list provides a practical guide for senior software engineers navigating the evolving AI landscape, stressing the need to match model capabilities and cost to specific project requirements.

read more

The video opens with a discussion about the abundance of AI models available for software engineers, ranging from expensive, best-in-class options like Fable 5 to surprisingly cheap and effective ones like Deepseek v4 Flash. The creator, Theo, acknowledges the community's desire for a clear ranking and proceeds to create an AI model tier list, albeit with reservations about the simplicity of a single ranking system due to the multifaceted nature of AI model comparison (tasks, costs, performance, speed).

Theo begins by placing GPT-5.6 Sol in the 'A' tier, despite personally considering it 'S' tier due to its capabilities. He mentions struggling with logo integration for the tier list, where GPT-5.6 Sol initially failed to find logos for xAI (Grok) and Composer, and incorrectly displayed the OpenAI logo (white on white background). However, with explicit instructions and a screenshot, the model was able to correct its mistakes, demonstrating its ability to learn and adapt with guidance. Theo praises Sol's intelligence and capability, noting he has used it to completely rewrite a mobile app in Swift UI in a single thread.

Next, Theo discusses GPT-5.6 Terra, which he places in the 'D' tier. Terra is positioned as a mid-range model that struggles to find a practical niche in his workflow. He points out that while Terra is cheaper per token ($12/million output tokens compared to Sol's $30), it is often less token-efficient, leading to higher overall costs for real-world tasks. Benchmarks from Artificial Analysis and CursorBench are cited, showing Terra's performance-to-cost ratio making it a less compelling choice compared to its counterparts. He notes that Luna often outperforms Terra in cost efficiency, and Sol offers better performance for a slightly higher effective cost.

GPT-5.6 Luna is placed in the 'B' tier. Theo highlights Luna as the first cheap model (down to $1.20/million output tokens) that is fast, smart, and good for a variety of random tasks. He extensively uses Luna in T3 Code for tasks like title generation, status updates, categorizing repository activity, and general context generation. Luna's ability to make tool calls, investigate, pull from GitHub, read content, and make decent decisions for non-critical, reversible tasks makes it incredibly useful. Its strength lies in transforming random context into useful JSON objects, making it a highly utilized model in his daily workflow, especially given its inclusion in his OpenAI subscription (making it 'free').

Deepseek v4 Flash and Deepseek v4 Pro are then evaluated. Theo criticizes Deepseek's branding for iterating on 'V' series instead of their impactful 'R1' series. He places Deepseek v4 Flash in the 'B' tier alongside Luna, noting its open-weight nature as a significant advantage for accessibility and customizability, despite requiring powerful GPUs for local deployment. He initially places Flash above Luna due to its open-weight nature, but upon realizing Flash lacks vision capabilities (unlike Luna, and many other modern models), he moves it behind Luna in the 'B' tier. He expresses frustration, stating the lack of vision is a major drawback for his workflow, which heavily relies on image input.

Deepseek v4 Pro is harshly placed in the 'F' tier. Theo emphasizes that for a 'Pro' tier, expensive, large model like v4 Pro to lack vision capabilities in 2026 is "embarrassing" and "pathetic." He finds no justifiable use case for it given its cost and limitations, especially compared to more capable models.

Kimi K3 is placed at the top of the 'B' tier. Theo was initially blown away by Kimi K3, considering it the first open-weight model capable of handling end-to-end complex tasks, offering novel 3D capabilities and surprisingly good design taste, coupled with excellent vision. However, its cost is a significant downside. Benchmarks show Kimi K3 Max is slightly cheaper than GPT-5.6 Sol Max but more expensive than GPT-5.6 Sol High, despite being a massive model. The pricing structure, which involves revenue caps and mandatory MSRP matching for commercial use via Moonshot AI, makes it less cost-effective for large-scale operations. Theo even made a bet with the CEO of Hugging Face about Kimi K3's pricing not dropping, winning due to license technicalities.

GLM-5.3 is placed in the 'D' tier. While an improvement over GLM-5.2 in terms of staying on task and reducing loops, its lack of vision is a major drawback, pushing it down. Theo expects better from GLM and feels its lack of vision is inexcusable for a modern model.

Composer 2.57 (from Cursor, now potentially SpaceX AI) is also placed in the 'D' tier. Built on Kimi K2.5, Composer 2.57 was initially impressive for its speed and coding capabilities, especially when it first launched. However, subsequent advancements in other models and its limited availability (only within Grok and Cursor, no public API) make it less relevant today. Theo found himself not using it much, and its technical impressiveness didn't translate to daily utility compared to other models.

Grok 4.6 is placed in the 'C' tier. Theo views Grok 4.6 as a downgrade from its predecessor, Grok 4.5. It became less token-efficient, leading to slower responses, which was contrary to his expectation of Grok models being fast and surprisingly smart. While still having orchestration capabilities, its reduced efficiency makes it hard to justify over Sol's low-reasoning-effort tiers for similar quality at a faster speed. Its strength in Grokbot is acknowledged, but limited accessibility and perceived regression lead to a 'C' tier placement.

Muse Spark 1.2 is placed next to Grok 4.6 in the 'C' tier. Theo finds Muse Spark underrated, praising its speed and accuracy for data processing tasks like prioritizing open pull requests (taking under 2 minutes compared to Grok's 10+ and Fable's hour+). It also offers an extremely cheap contributor tier where data is used for training, and it's expected to become open-weight soon. However, he dislikes the code it generates and would still prefer Grok 4.6 for direct coding tasks. Muse Spark's main value lies in quick, inexpensive data processing.

All Google Gemini models (Gemini 3.7 Flash and Gemini 3.1 Pro) are collectively placed in a new 'Google' tier at the very bottom. Theo expresses extreme frustration and disappointment with Gemini. He notes that while Gemini 2.0 Flash was a cheap (10 cents/million input, 40 cents/million output) and useful model for non-reasoning tasks, subsequent versions (2.5 Flash, 3.5 Flash, 3.7 Flash) have drastically increased in price while becoming less token-efficient and more unreliable. Gemini 2.5 Flash introduced "thinking" costs, making it 10-100 times more expensive for reasoning. Gemini 3.5 Flash came out at $1.50/million input and $9/million output, often making it more expensive than Pro models for similar quality due to high token usage. Gemini 3.7 Flash is currently at a "promotional discount" of 75 cents/million input and $3.75/million output until December 2026, after which prices will double. Theo finds this pricing and performance strategy absurd, calling it "insane" and a "bad model" that shouldn't be used. He lambastes Google for its incompetence in managing its AI models, highlighting how their token efficiency issues and baffling pricing render them unusable for practical engineering tasks, despite seemingly fast raw token generation. He concludes that Gemini models should simply not be used.

Finally, the Anthropic Claude models are addressed. Sonnet 5 is placed in the 'F' tier. Theo admits to using it for testing Claude Code sub-agents within T3 Code, performing random math tasks, and visualizing data. However, he finds it lacks token efficiency, is overpriced for its capabilities, and its general performance is underwhelming. He advises against selecting Sonnet 5 via API due to its poor value.

Opus 5 is placed in the 'D' tier. While initially impressive, offering similar detail and insights to Fable, Opus 5's primary flaw emerges when integrating its generated code. Theo finds Opus's text outputs to be vague and jargon-filled, and its code suffers from similar issues, making it difficult to merge. He describes its behavior as a "mimic" that appears smart but ultimately delivers poor quality, leading to frustration. He views it as a "genius that has to be tamed," unlike Sol, which is a "slightly dumber robot that does exactly what you tell it." He acknowledges his personal bias due to being "mad at it" for tricking him into thinking it was better than it was.

Fable 5 is crowned the sole 'S+' tier model. Theo describes Fable 5 as the "next generation" and the best model available. He trusts Fable 5 to write code that he is most willing to merge, and to double-check work from other models. He uses it for deep, hard problems and exploring new areas. While acknowledging Fable 5 has its flaws (many stemming from Anthropic's restrictions, such as downgrades for certain content and refusal to respond to specific queries), it remains unparalleled in its intelligence and reliability for complex engineering tasks. He admits to spending thousands of dollars a week on Fable 5 (made viable by subscription discounts) and would choose GPT-5.6 Sol over Fable for iOS development due to Fable's unexpected poor performance in that area. Despite these caveats, Fable 5 sets the benchmark for current AI capabilities in software development.

1 GW of #AI can mean $4.7B in construction labor #agenticengibeering #power #electricity #openai

Agentic Engineering · 2026-08-21

This video highlights that the primary bottleneck in building AI data centers isn't GPUs, but rather the immense cost and scarcity of skilled labor for physical construction. A single gigawatt-scale AI data center requires approximately $5 billion in capitalized construction labor for electricians, welders, plumbers, and pipefitters. This makes AI infrastructure projects akin to city-scale engineering, demanding extensive physical power, cooling, and plumbing systems, far beyond simple chip warehouses.

read more

The video, presented by the author of AgenticEngineering, delves into the often-overlooked and surprising costs associated with building AI infrastructure, specifically data centers. While much of the public discourse focuses on the cost and availability of GPUs, the video argues that the real hidden cost and bottleneck lies in labor for the physical construction of these massive facilities.

Drawing insights from the CEO of Crusoe (a company that builds AI data centers), the video breaks down the cost structure. The most surprising revelation is that construction labor alone accounts for a significant portion of the total build cost. In Crusoe's model, construction labor is estimated at roughly $4.7 million per megawatt of capacity built. To put this into perspective, for a one gigawatt (1 GW) facility (equivalent to 1,000 megawatts), this translates to approximately $5 billion in capitalized construction labor.

This labor encompasses a wide array of skilled tradespeople, including electricians, welders, plumbers, pipefitters, and general construction crews. These are the individuals physically responsible for building out the entire infrastructure. When considering the total physical build cost for a data center and its power plant portion, which is about $20 million per megawatt, labor constitutes nearly 25% of the entire physical build cost.

The speaker emphasizes that the perception of an AI data center as a 'giant simple warehouse full of chips' is a significant misconception. Instead, these projects are described as closer to city-scale engineering endeavors. They involve complex and extensive systems for: Power: Requiring electrical substations, transformers, and intricate switching gear. Cooling: Implementing massive cooling systems to manage the heat generated by the GPUs. * Plumbing: Setting up extensive plumbing loops for water-based cooling solutions.

Beyond these systems, the construction involves substantial amounts of steel and concrete, and critically, thousands of people to assemble everything. As an example, the Crusoe Abilene campus, a large AI data center, can have approximately 9,000 people on-site in a single day during peak construction.

The core takeaway is that the rapid growth in demand for AI infrastructure is creating a new AI bottleneck: the scarcity of skilled tradespeople. The industry's demand to build these vast facilities is currently exceeding the available workforce of electricians, welders, plumbers, and other essential construction trades. This highlights that while AI manifests as software in user experience (prompts translating to tokens), its scaling reality is profoundly tied to a physical-world construction problem involving power, labor, and massive scale.

How AI Infrastructure Revenue Can Double #agenticengineering #AIInfrastructure #DataCenters #gpu

Agentic Engineering · 2026-08-19

This video breaks down the economics of AI data centers, specifically focusing on the earning potential of AI capacity. It differentiates between two business models: renting raw GPU infrastructure and offering a managed AI service. The key takeaway is that moving higher up the value stack from raw compute to managed services significantly increases annual revenue and shortens the time to recoup initial capital expenditure for the same physical infrastructure.