Level up your AI understanding with The Neuron’s July 2025 AI Research Digest.
This month’s issue gathers the most mind-bending, paradigm-shifting discoveries from across the AI research landscape—so you don’t have to scroll through dozens of papers to find the gems.
Each month, our team surfaces the “holy smokes, did you see this?” moments—those breakthroughs that spark heated debates in our Slack, trigger existential engineering crises, and reshape how we think about the future. July’s collection is no exception.
July 16
- Exponential view explains why Kimi K2 is the AI model that should worry Silicon Valley.
- Check out OpenAI’s new podcast with OpenAI COO Brad Lightcap and Chief Economist Ronnie Chatterji that covers AI and jobs.
- METR's analysis of 9 AI benchmarks revealed that human task completion times vary by over 100x—from 2 minutes for computer use to 1+ hours for reasoning tasks—explaining why AI models are advancing unevenly across domains with median improvement cycles of 4 months.
- A pretty well-covered study showed AI actually slowed down developers; one of those developers tells his tale in this eye-opening blog recap.
- Very interesting: Calvin French-Owen reflected on his 14-month tenure at OpenAI, describing a rapidly expanding organization that grew from 1K to 3K employees with an environment resembling “the early days of Los Alamos.”
- Spyglass writes that we’re still in the “throw money at it” era of the AI boom supercycle.
- A new publicly developed large language model trained on the “Alps” supercomputer will be released in late summer 2025, featuring fluency in over 1,000 languages and complete open access.
- Is AI perma-bear Gary Marcus right that at some point, AI engineers will have to embrace neurosymbolic AI? Or is the bitter lesson that reinforcement learning will always eventually win out in AI training destined to debunk this?
- Andrej Karpathy thinks RL is not the whole story, and Yuchen Jin of Hyperbolic proposed a lesson-based learning system with more feedback loops.
- Gwern writes that perhaps what today’s AI is missing before it can create novel discoveries is a day-dream loop
- There’s a debate about “LLM inevitabilism,” and how this rhetorical framing strategically advantages tech leaders by narrowing debate over if something should be allow to happen to questions of adaptation as opposed to if it should happen at all.
- Leading AI researchers from OpenAI, Anthropic, Google DeepMind signed a position paper urging companies to monitor advanced models' internal reasoning through Chain-of-Thought monitoring.
- Here’s how to vibe code and actually ship, y’know, real code.
July 11
- In a survey of 1K+ C-suite execs (CEO, CFO, etc), Icertis found 90% claim increased business pressure, with 46% struggling to show AI ROI and 46% racing to keep up with AI innovation. Executives cited AI washing (51%) and unclear AI use cases (44%) as investment decision barriers, while 90% expected tariffs to hurt profits.
- Researchers designed an AI that can “autonomously discover and exploit vulnerabilities” in smart contracts—put another way, it can steal crypto (paper).
- AI is increasingly making Apple look like a loser to the stock market, with the stock down 15% due to concerns over tariffs and its ability to compete with AI, and some analysts believe acquiring Perplexity is its only shot at keeping up.
- A basic security flaw in the McDonalds hiring AI called “McHire” (built by Paradox) exposed millions of applicant’s sensitive data to hackers. Why? Because of the password “123456”.
- Check out this fascinating argument that language models like ChatGPT are actually the perfect validation of 1980s French Theory—they prove language can generate meaning purely through statistical relationships between signs, without any understanding or consciousness, exactly as structuralists claimed before anyone imagined such machines could exist.
July 9
- AI is driving massive data industry consolidation, where companies are abandoning fragmented tools for unified platforms, as shown by Databricks' $1B Neon acquisition and Salesforce's $8B Informatica deal.
- The creator of Soundslice built an ASCII tab import feature simply because ChatGPT was incorrectly telling users it existed, revealing how AI hallucinations might increasingly shape product roadmaps as users trust AI-generated misinformation.
- Ethan Mollick published a new essay “against brain damage” (and misinterpretations of the MIT study), defending the use of AI in education.
- Fireship explained the Soham Parekh “overemployed” debacle (the developer who worked for at least 4 startups in San Francisco at once) and how its a symptom of the larger job market’s sickness.
- Sakana AI revealed a new breakthrough of letting different language models work as a "dream team" through its TreeQuest framework, improving performance by 30% while demonstrating how AI systems achieve collective superintelligence.
- This AgentCompany benchmark reveals current AI agents can autonomously handle up to 30% of simulated workplace tasks—suggesting we should focus on task-level automation rather than assuming entire professions will be automated (for now).
- The Algorithmic Bridge gave a great AI industry critique warning about growing dependency on fundamentally flawed AI tools, while questioning whether the industry prioritizes solving core problems like hallucinations or simply races ahead.
July 2
- If you watch one AI interview this week, watch this one; It’s Matt Berman’s chat w/ Dylan Patel, the best writer covering AI chips and data centers, and who is a wealth of insight on what’s happening in the industry. For example:
- All the complicated in’s and out’s of Microsoft and OpenAI’s bumpy relationship right now (15:33).
- A pretty technical but very informative post-mortem on OpenAI’s GPT 4.5 and what went wrong (22:28)
- The beef between Apple and NVIDIA, which could be one reason why Apple is behind in the AI race (31:06).
- NVIDIA’s role in the AI industry as both a cloud kingmaker (41:31) and cloud disruptor (43:13)—as Dylan says, “You don’t mess with God. What Jensen giveth, Jensen taketh,” and cloud companies are mad.
- Elon’s Grok 3.5 and what’s going on there (52:49).
- What model he goes to the most (51:30), which depending on the topic is ChatGPT o3, Claude 4, or Gemini 2.5 Pro, and Grok 3. You should use the best model for you, but Dylan’s pretty based, so worth considering his use cases for each!
- And who his pick is to win the “superintelligence” race… and why (1:01:01).
