Self-adapting models & ChatGPT in Barbies
Our weekly updates just got bigger! Free subscribers will continue to get the top stories each week, while Paid subscribers will get a few extra bullets each week to keep you even more up-to-date.
last week’s top stories
🧠 MIT researchers enable self-adapting language models. MIT’s SEAL framework lets a deployed LLM rewrite its own weights for new tasks without full retraining. A policy module judges each weight patch and rolls back regressions. Translation and summarization scores jump 12% while compute falls 30%. Read more
💰 Meta buys Scale AI for $14.8 billion. Meta will fold Scale’s labeling and eval pipelines into Llama pre-training and agent benchmarking. CEO Alexandr Wang joins Meta’s super-intelligence group, reporting to Yann LeCun. Closing awaits FTC review and a Q4 capex carve-out. Read more
📺 Kalshi airs $2k AI-generated NBA Finals spot. The 30-second clip was built with Google Veo 3 and cloned narration in under three days. Total spend: cloud credits, one editor, zero cameras. The alligator-riding banker became instant meme fuel. Read more
🧸 Mattel taps ChatGPT Enterprise for toy design. A multiyear deal gives product teams private GPTs for concept testing, supplier search, and marketing copy. OpenAI and Mattel will co-build a streaming “AI dollhouse” platform for connected toys. Rollout starts with Barbie digital playsets. Read more
⚖️ Disney and Universal sue Midjourney over training data. Studios allege unauthorized ingestion of storyboards and stills in Midjourney’s diffusion corpus. The filing seeks data deletion plus triple damages under the DMCA. First case-management conference lands August 4. Read more
🌐 Dia browser beta blends tabs with agents. The Browser Company opened invites for Dia, a Chromium fork with an embedded 7B LLM and plugin graph. A sidebar parses page context, drafts replies, and schedules follow-ups. Token spend caps at 5 M per day per user. Read more
💸 OpenAI drops o3 prices 80 % and adds o3-pro tier. Input now runs $2/M tokens, output $8; o3-pro offers 8x throughput plus expanded vision context. Read more
🍎 Apple opens on-device “Apple Intelligence” SDK. WWDC25 delivered foundation models that run inside A18 and M4 secure enclaves. Live translation, call screening, and semantic search pipe through a new API. Xcode 17 beta ships today.
🧪 AI Research of the Week
Understanding the Impacts of Generative AI Use on Children
From the Alan Turing Institute's Children & AI team, backed by the LEGO Foundation
Jake’s Take: This paper explores UK-based surveys of 780 children and 1,001 teachers, alongside hands-on school workshops. Turns out, one in five kids aged 8–12 actively uses ChatGPT, despite age restrictions intended to exclude them. Private-school kids are significantly ahead of peers in state schools regarding AI exposure. Children voiced sharp observations on bias, misinformation, and the environmental costs of generative AI, and notably highlighted AI’s frequent failures in representing their identities through images. Parents are more worried about toxic content than cheating, while teachers appreciate AI as prep support yet remain skeptical about its impact on genuine student participation.
Key recommendations here emphasize kid-friendly protections, data rights, and direct youth involvement in shaping policies. Industry reluctance to address these concerns head-on invites inevitable criticism (and legal headaches).
and then, even more news…
🌪️ DeepMind model sharpens hurricane forecasts. A transformer-U-net stack ingests satellite and buoy feeds, outputting 5 km-grid ensemble tracks. Mean absolute error drops 38 km at 72-hour lead time. NHC pilot starts July. Read more
🛠️ GitHub Copilot obeys repo-level instructions. A .github/copilot-instructions.md file now guides code-review suggestions the same way it shapes chat. Teams can enforce style and licensing rules across pull requests. Feature is live for all paid seats. Read more
🛡️ Microsoft readies classified Copilot for the Pentagon. Microsoft 365 Copilot is being hardened for GCC High with air-gapped clusters, aiming for ATO by summer 2025. The deal could cover a million seats and replace VBA macros. DISA pilots begin this fall. Read more
🏭 Oracle and Nvidia unveil turnkey “AI factories.” OCI will host GB200 NVL72 racks plus NeMo microservices under a per-token plan. A full rack rents for $4M per year with Kubernetes NIM orchestration. NAB is first customer. Read more
🌍 Nvidia’s “Climate in a Bottle” boosts Earth-2. The model simulates global weather at five-kilometer grids, shrinking century-scale runs from weeks to hours on Blackwell GPUs. Turing Institute feeds the output to insurers through an open NetCDF API. Read more
🧩 Databricks unveils Agent Bricks and MLflow 3.0. Mosaic AI auto-generates function-calling chains and logs agent telemetry to Delta tables. MLflow 3 tracks latency histograms and guardrail incidents across endpoints. GA lands in workspace version 17.3. Read more
🧮 Mistral rolls out Magistral reasoning model. The 24B-parameter Magistral Small scores 70.7% on AIME2024 and ships under Apache-2. Magistral Medium hits 73.6% with 40k context and chain-of-thought voting. Both live on Hugging Face and in Mistral’s API. Read more
🚀 Amplitude pairs Amazon Bedrock agents with growth analytics. The reference stack links Bedrock multi-agent orchestration to Amplitude event streams, surfacing retention drivers from natural-language prompts. Sample code runs with Guardrails disabled. Read more
🔑 Databricks AI Gateway hits GA. Gateway centralizes credential rotation, rate limits, and spend tracking for external model endpoints. Policies sync with Unity Catalog and raise rogue-token alerts in real time. Terraform provider ships in the GA bundle. Read more
📈 Nvidia posts data-flywheel blueprint for agent retraining. The workflow captures user logs, ranks samples with NeMo evaluator, and runs LoRA fine-tunes via Argo pipelines. A retail chatbot gained seven accuracy points after two loops. Read more
📊 Oracle lifts FY 2026 outlook on AI demand. Cloud growth guidance rises to 40 % as GPU workloads pile into OCI and gen-AI tools bundle into SaaS contracts. Q4 revenue hit $15.9 B, topping estimates by $300 M and sending shares up 7%. Reuters