Welcome to August 4, 2026
The Singularity has begun filing its own optimization tickets. Asari AI’s self-improving “co-inventor” agents rebuilt the inference stack for DeepSeek v4 Pro and GLM 5.2 on B200s, lifting throughput and interactivity up to 16%. Intology’s Locus agent leads PostTrainBench, post-training models unsupervised in ten H100-hours and beating human tuners on the harder variant. On MirrorCode, which tests whether agents can rebuild whole software projects from scratch and pass every test, Claude Fable 5 solves 64% to GPT-5.6 Sol’s 20%. Gemini 3.5 Pro reportedly lands next week as “a solid model, but it won’t clearly overtake the best models from Anthropic or OpenAI,” which is what a plateau looks like from inside a vertical climb. MiniMax-H3 runs text, image, video, and audio through a single 33B-parameter transformer to generate video with native 32-kHz stereo at 2K, and an investor one-shotted Jim and Dwight debating coding agents. The Office now writes its own episodes about what replaced the office.
Mathematics is the cleanest gauge. arXiv math uploads are spiking while budgets flatlined, the clearest signal of AI influence. VibeMathed supplies receipts: 427 problems tracked, 312 resolved, 205 in July alone, up 193% over June, a third checked in Lean. Wet labs run the same loop with slower atoms. Tianjin’s REAP pairs a hybrid-loss model with robotic experiments for a 57-fold activity gain in cytochrome P450 BM3 in five cycles, 104-fold in Sortase A. Then it closes on us. Science Corp’s SciFi headstage, from $2,048, moves 2.5 Gbps across thousands of channels with on-device models, so feedback never leaves the skull.
Code is being deprecated by its own output. Elon Musk argues “Source code is on the verge of becoming like assembly,” with AI emitting the binary directly. One founder notes SaaS multiples fell from 18x revenue to 3.4x because “the difficulty of building software is TRENDING TO ZERO,” leaving distribution as the one asset a model cannot clone. The tools eat themselves next. OpenAI’s Thibault Sottiaux calls Codex a good harness that “will seem primitive in 2-3 months,” since “the next generation of models need more than your laptop.”
What they need is continental machinery. Caterpillar posted record $20.5 billion sales, data-center power generation up 29%. Anthropic signed a $10 billion deal with Nvidia-backed Volta for 133 MW of hydropowered Vera Rubin capacity in Norway, while Google built one of history’s largest financing programs to route $150B of chips to the same lab. The racks are leaving the planet. SpaceX is partnering with Nvidia to fly those same Rubin GPUs and Vera CPUs on Starmind AI1 satellites, “datacenter class space compute.” DeepMind’s chief strategy officer says what justifies this capex is recursive self-improvement, AI building better AI. On the ground, SpaceX prepaid Grimes County $10 million on a Terafab deal that could reach $119 billion. PC makers began buying Chinese CXMT DRAM. Huawei’s Liao Heng warned that Western die and HBM scaling is nearing a physical wall, pitching “Tau Scaling Law” instead. Washington’s reply is an FCC-drafted ban on Chinese data center gear.
Policy oscillates faster than the hardware. Officials weighed sanctions and blacklists against open-source Chinese labs, then reversed after Jensen Huang lobbied against restrictions OpenAI and Anthropic wanted, picking competitiveness over containment. Lab staffers reviewed the finished voluntary evaluation framework, still undisclosed because unclassified “doesn’t mean we are going to broadcast them to everyone.” Beijing frets that Mythos is an offensive cyber weapon ahead of Xi’s September visit. Palantir’s Alex Karp accused the labs of “trying to drug addict us to a future they believe they control,” and posted 93% revenue growth to $1.94 billion. Amazon crossed $3 trillion on its fastest AWS growth since 2021. As researchers keep hopping between labs, Dario Amodei reportedly worries that new talent now comes to Anthropic for the money rather than the mission.
Legacy institutions are being re-audited. Tax preparers use AI under 2013 guidance. UNAM, Mexico’s largest university, ran its first remote entrance exam and got scores so implausible that 58,000 applicants must sit it again, the proctoring AI having lost to the test-taking AI. Mariana Minerals raised $310 million after restarting an idled Utah copper mine in four months on autonomous software, since electrified intelligence eats metal. Finance re-embraced blockchain, stablecoins at $300B and tokenized funds quadrupling, even as the IMF warns failures “can propagate faster than institutions or supervisors can respond.”
A far older ledger may open next. A presidential speech confirming that some UAP are of non-human origin is under consideration before November, following an August 1 memo that freed employees and contractors from NDAs.
We do these things not because they are easy, but because they soon will be.



Love this: We do these things not because they are easy, but because they soon will be.
thanks for a great post again, and for the Kennedy misquote. My AI reads all your posts, highlights what is most important to me and provides background for me on concepts I may not be familiar with. Does the same with Moonshot transcripts. It knows me from thousands of interactions on work and play over the past year. There is no benchmark I have seen on how rapidly context memory is growing but it's rapidly becoming of enormouse value to me. Youtube traffic on this is low. Maybe a slot about this on Moonshots?