🎧 Listen Ads-Free: Subscribe to DjamgaMind via Apple Podcasts for a pure, ad-free experience at https://podcasts.apple.com/us/channel/djamgamind/id6760446113

#DJAMGAMIND #AIUNRAVELED

Summary: In today’s briefing, we analyze the critical divergence between AI capability and reliability. We deconstruct the launch of OpenAI’s GPT-5.5, which dominates reasoning benchmarks but struggles with severe hallucinations compared to Claude Opus. We explore the environmental fallout as Alphabet, Amazon, Meta, and Microsoft abandon net-zero pledges in favor of natural gas to power AI data centers. We also cover Moonshot AI’s Kimi K2.6 executing multi-day autonomous coding, Tim Cook’s warning regarding soaring memory chip costs (”RAMageddon”), Elon Musk’s courtroom admission about distilling OpenAI’s models, and a severe Linux vulnerability catching the world flat-footed.

Important Topics:

• The GPT-5.5 Paradox: OpenAI releases GPT-5.5, topping the Artificial Analysis Intelligence Index but exhibiting a significantly higher hallucination rate (85.5%) compared to Claude Opus 4.7 (36.1%).

• Big Tech Climate Collapse: Alphabet, Amazon, Meta, and Microsoft are reverting to natural gas power plants and abandoning net-zero timelines to meet the massive energy demands of AI data centers.

• Kimi K2.6 Open-Weights Autonomy: Moonshot AI updates Kimi K2.6, a 1-trillion parameter model capable of instantiating up to 300 parallel sub-agents for multi-day autonomous coding tasks.

• Apple’s “RAMageddon”: Apple posts a record $111.2B quarter, but CEO Tim Cook warns that AI industry demand has quadrupled memory chip prices, severely constraining the supply chain.

• Musk Admits to Model Distillation: In federal court, Elon Musk admits that xAI used distillation techniques on OpenAI’s models to train Grok.

• Severe Linux Vulnerability: A zero-day flaw named “CopyFail” (CVE-2026-31431) goes public, allowing attackers full control over unpatched Linux servers and Kubernetes containers.

• Meta Fires 1,100 AI Trainers: Meta contractor Sama fires over a thousand Kenyan workers after whistleblowers reveal they were reviewing sensitive Ray-Ban smart glasses footage.

• Strategic Thinking in LLMs: Researchers prove that Gemini and GPT models use more sophisticated, sequential strategic planning in games (like Rock-Paper-Scissors) than human players.

🛠️ The AI Executive Toolkit: Deploy real infrastructure. Get the hand-picked, forensic-vetted implementation stack built for the C-Suite at https://DjamgaMind.com/Toolkit.

• Google Workspace: Professionalize your firm’s infrastructure with secure, cloud-based collaboration and branded communication: https://referworkspace.app.goo.gl/Q371

⚗️ PRODUCTION NOTE: We Practice What We Preach.

AI Unraveled is produced using a hybrid “Human-in-the-Loop” workflow.

GPT-5.5 Outperforms, Hallucinates

The latest update of OpenAI’s flagship model sets new states of the art in important benchmarks but has difficulty distinguishing between what it does and doesn’t know.

What’s new: GPT-5.5 is a closed vision-language models that’s built for agentic coding, computer use, and knowledge work. GPT-5.5 Pro is the same model but processes reasoning tokens in parallel during inference. OpenAI set the API prices at roughly double the per-token rates of GPT-5.4.

• Input/output: Text and images in (up to 1 million tokens via API, 400,000 tokens in Codex), text out (up to 128,000 tokens)

• Features: Five levels of reasoning (xhigh, high, medium, low, none), tool use, web search, structured outputs, tool search (API only, loads tools on demand rather than all at once), Fast mode (Codex only, generates tokens 1.5 times faster at 2.5 times the price)

• Performance: Tops Artificial Analysis Intelligence Index and ARC-AGI-2

• Availability/price: GPT-5.5 available in ChatGPT with Plus, Pro, Business, or Enterprise subscription and in Codex for those tiers plus Edu and Go; GPT-5.5 Pro available in ChatGPT with Pro, Business, or Enterprise subscription: GPT-5.5 API $5/$0.50/$30 per million tokens of input/cached/output, GPT-5.5 Pro API $30/$180 per million tokens of input/output with no cached discount

• Undisclosed: Architecture, parameter count, training data and methods

How it works: OpenAI disclosed few details about how it built GPT-5.5. As is typical of high-performance models, the training data was a mix of publicly available data scraped from the web, licensed from partners, and collected from users and human trainers. The model was trained via reinforcement learning to reason before responding.

Big AI’s Plans Strain CO₂ Pledges

Commitments by large AI companies to limit emissions of greenhouse gases are at risk as those companies pursue a massive build-out of data centers, many of which will be powered by fossil fuels in the near term and possibly beyond.

What’s new: Alphabet, Amazon, Meta, and Microsoft have begun to acknowledge that keeping up with projected demand for AI is interfering with earlier plans to stop raising the concentration of greenhouse gases to the atmosphere, Associated Press reported. (Disclaimer: Andrew Ng is a member of Amazon’s board of directors.)

How it works: Electricity consumed by top tech companies has increased significantly over the last few years, and with it their emissions of greenhouse gases that contribute to climate change, despite ongoing efforts to reduce emissions. While they have emphasized clean sources of energy including wind, solar, geothermal, and nuclear, lately they have begun to develop natural gas power plants to meet rapidly rising demand for AI.

Kimi K2.6 Challenges Open-Weights Champs

Moonshot AI’s updated Kimi model handles longer autonomous coding sessions and scales up its multi-agent orchestration relative to its predecessor.

What’s new: Kimi K2.6 is a 1 trillion-parameter vision-language model that performs neck and neck with Qwen3.6 Max Preview and the newly released DeepSeek V4 and falls just behind top closed models. It’s designed to generate code in a plan-write-test-debug loop that can last for days, and it can instantiate hundreds of agents that collaborate on a single task. It also produces fewer hallucinations than its predecessor.

• Input/output: Text, images, and video in (up to 256,000 tokens), text out (up to 98,000 tokens)

• Architecture: Mixture-of-experts, 1 trillion parameters total, 32 billion active per token, MoonViT vision encoder

• Features: Tool use, web search, native INT4 quantization, “preserve thinking” mode, agent swarm

• Performance: Tops other open-weights models on the Artificial Analysis Intelligence Index but trails leading proprietary models

• Availability/price: Weights free to download from Hugging Face under a modified MIT license that permits commercial uses with attribution for products with more than 100 million monthly active users or more than $20 million in monthly revenue, free chat interface at kimi.com and Kimi mobile app, API access via Moonshot $0.95/$0.16/$4.00 per million input/cached/output tokens

• Undisclosed: Training data and methods

How it works: Kimi K2.6 reuses the architecture introduced with Kimi K2 and refined in Kimi K2.5, including the multi-headed latent attention (an attention variant that reduces memory requirements by compressing keys and values) and MoonViT vision encoder (400 million parameters). Moonshot has not disclosed how Kimi K2.6 differs with respect to training data and methods.

Strategic Thinking in LLMs vs. Humans

While large language models can behave in human-like ways, the similarities are superficial. A simple strategy game revealed clear differences in their strategic approaches.

What’s new: Caroline Wang and colleagues at University of Texas at Austin and Google interpreted patterns of decision-making by humans and LLMs as they played the classic game of rock-paper-scissors. They found that LLMs sometimes model their opponents with greater sophistication than people do.

Key insight: Given recorded gameplay, an LLM can iteratively improve code that predicts a player’s next move. If the code predicts the player’s actions with significant accuracy, we can assume that its decision-making algorithms are functionally similar to those the player used. Computer code is interpretable, making it possible to discern such algorithms and compare those used by humans and LLMs.

How it works: In games of rock-paper-scissors, he authors pitted individual LLMs (Gemini 2.5 Pro, Gemini 2.5 Flash, GPT-5.1, and GPT-OSS 120B) against each of 15 preprogrammed bots of varying complexity. They recorded each player’s moves in 20 games of 300 sequential rounds each. Previous work provided records of similar records of games between humans and the same bots. The authors tracked the round-by-round choices made by each player — AI and human — and whether they won, lost, or tied. Then they used AlphaEvolve, an agentic method that iteratively optimizes code through an evolutionary process, to improve Python programs that predicted the next move for each LLM individually and humans as a group.

• AlphaEvolve initially processed the game data using a simple template program written by the authors. In each of an undisclosed number of evolutionary steps, Gemini 2.5 Flash proposed modifications to improve a function that balanced simplicity (as measured by Halstead effort) and evaluation likelihood (how well a program predicted a player’s choices).

• For each player, the authors selected the simplest program that achieved near-maximum predictive accuracy within a small margin from the best. Each program produced the best evaluation likelihood (higher is better) for the player it had evolved to predict. That is, it represented its corresponding player’s behavior better than that of any other player.

Results: Using game data that AlphaEvolve didn’t process, the authors compared how well each program predicted the other players’ moves. Then they examined the programs to determine what strategies each player used.

Why it matters: While researchers have found ways to understand some aspects of neural network behavior, large language models remain black boxes in many ways. Synthesizing code directly from LLM behavior offers a powerful tool to interpret their decision-making.

We’re thinking: It’s tempting to assume that LLMs learn to mimic human behavior as represented by their training data. Finding that they can encode a gaming strategy more systematically than the average human demonstrates a different sort of learning.

Apple hits record sales despite chip shortage

• Apple brought in $57 billion from iPhone sales during its record March quarter, contributing to total revenue of $111.2 billion, with Tim Cook crediting strong demand for the iPhone 17 lineup on Thursday’s earnings call.

• Cook cautioned that memory chip costs will climb significantly starting in June due to “RAMageddon,” the AI industry’s heavy appetite for memory chips, which has already quadrupled RAM prices and may push iPhone prices higher.

• Apple’s March spending on memory chips already rose, though the company offset costs by selling stockpiled inventory, and Cook told Reuters there is “just a little less flexibility in the supply chain at the moment for getting more parts.”

Musk says xAI trained Grok on OpenAI

• Elon Musk admitted on the stand in a California federal court on Thursday that xAI trained Grok using distillation on OpenAI models, saying the practice was common across AI companies when asked directly.

• The admission came during Musk’s trial against OpenAI, Sam Altman, and Greg Brockman, where he alleges they broke the original nonprofit mission by shifting the entity to a for-profit structure.

• Musk also ranked the leading AI providers during testimony, placing Anthropic first, followed by OpenAI, Google, and Chinese open source models, and described xAI as a smaller company with just a few hundred employees.

Tesla starts Semi truck mass production after 9 years

• Tesla has kicked off mass production of its Semi electric truck nearly a decade after the 2017 unveiling, with the first big rig rolling off the high volume line at a dedicated plant near Gigafactory Nevada.

• The Semi comes in a 325-mile Standard Range trim priced around $260,000 and a 500-mile Long Range version near $300,000, both packing a 1,072HP tri-motor system that charges at up to 1.2MW on Megachargers.

• Deliveries start later this year and undercut the Freightliner eCascadia ($400,000, 230 miles) and Volvo VNR Electric ($350,000, 275 miles), though Tesla won’t hit the factory’s 50,000-truck yearly peak output.

Severe Linux threat catches world flat-footed

• Researchers at security firm Theori released attack code on Wednesday for a Linux bug called CopyFail that lets regular users seize full control of almost every version of Linux, leaving companies racing to protect servers and personal machines.

• Known as CVE-2026-31431, the flaw was quietly reported to Linux kernel maintainers five weeks ago and fixed in updates such as 7.0 and 6.19.12, but most Linux distributions had not yet shipped those fixes when the code went public.

• One script works against every unpatched system without changes, letting attackers take over shared servers, escape Kubernetes containers that isolate apps, and sneak the code into pull requests so it runs inside automated build and deployment pipelines.

Meta fires 1,100 AI trainers over Ray-Ban leaks

• Meta has cut ties with Sama, a Kenya-based contractor that trained its generative AI systems using Ray-Ban smart glasses footage, triggering the layoff of 1,108 workers after some spoke out about the recordings they reviewed.

• Sama employees told Swedish newspapers in February that they labeled footage showing banking information, private conversations, naked people in bathrooms, and intimate encounters, often captured from subjects who seemingly did not know they were being recorded.

• Meta says its terms of service cover these details and the glasses need explicit permission to engage AI mode, but Sama workers reported being forced to sit idle under tighter security as the firm hunts for the whistleblowers.

1X opens US humanoid factory targeting 100,000 NEO robots

• 1X has begun full-scale production of its NEO humanoid robot at a new 58,000-square-foot factory in Hayward, California, with plans to build more than 100,000 units per year by 2027.

• The factory uses a vertically integrated model, with 1X designing and making motors, batteries, sensors, structures, and transmission systems in-house, and its first-year run of over 10,000 units sold out within five days of its October launch.

• Each NEO runs on NVIDIA’s Jetson Thor platform for onboard AI inference and is trained using NVIDIA Isaac simulation tools, with customer shipments starting in 2026 through a $20,000 early access program or a $499 monthly subscription.

Mine hunting:

The US Navy is turning to Domino, a startup that helps companies build, deploy, and manage AI to find mines in the Strait of Hormuz. With a fresh $100 million contract, Domino will help the American military “make underwater mine detection faster, more accurate, and less dependent on human sailors.” The Navy-Domino partnership underscores the rising importance of AI in military contexts, including keeping global shipping lanes free from explosives.

Stablecoins are big business:

Polymarket’s rapid growth was partly enabled by Fun, a startup that provides crypto and fiat on- and off-ramps for customers. The fintech upstart now processes $18 billion in annual payment volume, enough for venture capitalists to put $72 million into its coffers as part of an outsized Series A. As the stablecoin market grows, so too will demand for financial technology that makes their use simple and safe.

The race to build efficient AI accelerates:

Nebius, a public neocloud, announced today that it plans to purchase Eigen AI. Eigen helps make AI inference and certain AI training activities more efficient. Given that Nebius serves AI compute demand today and there’s not enough to go around, snapping up a smaller company to make existing GPUs stretch further makes good sense. And if you are Nebius and want to defend your valuation, then working to ensure your gross margins are as attractive as possible is simply good business.

The White House’s Anthropic stance gets complicated

The Rundown: The White House is pushing back on Anthropic’s plan to more than double the private sector’s access to its Mythos AI over compute concerns for its own use, just as a national security memo prepares to address parts of the Pentagon feud.

The details:

• Anthropic wanted access expanded from about 50 firms to nearly 120, with U.S. officials citing compute strains that could impact government use.

• A White House AI memo will reportedly push multi-vendor AI adoption for agencies and address some of Anthropic’s worries that led to the initial feud.

• Axios reported that the government action would “allow agencies to get around the supply chain risk designation”, despite the current legal battle.

• GPT 5.5 reached similar cyber capabilities to Mythos, with former AI czar David Sacks saying all frontier models will reach the level in 6 months.

Why it matters: The White House is changing its tune on Anthropic, seemingly largely in part to wanting more access of its own to the powerful Mythos. But with Sec. of War Pete Hegseth saying Thursday that Anthropic is “run by an ideological lunatic”, there is some internal division between wanting to bury the hatchet vs. continuing the fight.

Gemini moves into Google-powered cars

Google is beginning its Gemini upgrade for vehicles with Google built-in, swapping out Assistant for a more conversational system that handles navigation, messages, music, vehicle questions, and hands-free controls across compatible cars.

The details:

• Drivers can ask for changes to car settings like temperature, control the radio, and pull from Google Maps for customized updates or route planning.

• A beta Gemini Live mode supports conversations for learning and brainstorming, with Gmail, Calendar, and Home integrations coming later.

• Gemini can also pull vehicle-specific answers from manufacturer manuals for car assistance and battery status or charging stations for EV cars.

• The rollout comes to compatible cars in the U.S. first, with General Motors also announcing the feature for ~4M of its vehicles from model year 2022 onward.

Why it matters: One day, AI integrations in cars will be as common as a radio (and eventually the systems will all be driving the cars, too) — but for now, we’re still in the infancy of the rollout. These initial features are fairly basic, but a step on the path towards ‘smart car’ systems of the AI age that provide a serious intelligence upgrade.

OpenAI finds source of ChatGPT’s goblin obsession

Image source: OpenAI

OpenAI just traced ChatGPT’s habit of peppering its responses with goblins, gremlins, and assorted fantasy creatures back to a single reward signal in its ‘Nerdy’ personality, which ended up bleeding into model behavior throughout releases.

The details:

• After ChatGPT-5.1’s November launch, ‘goblin’ mentions jumped 175% in user conversations, with ‘gremlin’ up 52% and other creatures seeing similar spikes.

• When OpenAI mapped creature use across personalities, the Nerdy preset lit up, driving two-thirds of all goblin mentions from just 2.5% of traffic.

• Even users who skipped Nerdy got goblins, with fine-tuning loops recycling the creature-favored outputs back into ChatGPT’s default mode.

• OpenAI retired Nerdy in March and shipped GPT-5.5 with a Codex prompt specifically banning goblins, gremlins, ogres, trolls, raccoons, and pigeons.

Why it matters: ChatGPT’s goblin-mode is a fun little quirk for your Friday, and another example of how weird LLMs can truly be. A reward in a single personality mode led to a pattern of creature preferences that trickled across chats around the globe. Just like Anthropic’s Golden Gate Claude, we might need a standalone GoblinGPT.