A Frontier Model Edits Another Model's Weights as Anthropic Discloses $895 Billion in Compute Commitments

AI RESEARCH: A FRONTIER MODEL EDITS ANOTHER MODEL’S WEIGHTS

On November 2, Rikiya Takehi, a first-year PhD student at MIT CSAIL, announced work he calls Weight-Space Programming. He wrote that GPT-6 Astra was asked to improve the math ability of Qwen3.5-0.8B, a small open model, and did so by changing only 0.016% of its weights, with no gradients and no training. [1]

The blog post, by Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab of MIT CSAIL, reports that Astra traced the small model’s reasoning loops to one direction written by two of its layers. It removed that direction from 131,072 weights. With 50 forward passes allowed, MATH-500 rose from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and the instruction-following test IFEval from 21.0% to 62.3%. [2]

The authors also report that Astra named which game a small network was trained to play from its 17,253 raw weights, fixed a failing car-control network by editing 3.5% of its weights in one try, and wrote game-playing networks from zero weights that beat networks trained with PPO for 100 million steps. [2]

They list limits. Astra could not tell an honest Gemma 3 270M model from a sycophantic copy using weights alone, and it could not raise the math score of a Gemma model that did not loop. They write that editing weights “could remove an obstacle that hid an ability, but it could not add an ability the model lacked.” They say they are building a benchmark and harness, and they released full logs. [2]

On November 3, Sean Goedecke published a rebuttal on his blog. He argued that the game result follows from first-layer weight norms that three lines of Python can compute, and that the Qwen fix matches the known steering and abliteration method a competent human researcher would use. He compared direct weight editing to tweaking compiled binaries, and said training remains the better way to improve models. [3]

The same day, Jacob Lauritzen, chief technology officer of the legal AI company Legora, wrote that he had a model he called “fable” hand-write the weights of a 100,000-parameter network that tells real legal citations from fake ones. He said it “basically one-shotted a full NN that’s better than training it.” The post did not identify the model further. [4]

COMPUTE: ANTHROPIC’S $895 BILLION IN COMMITMENTS

Anthropic, the maker of Claude, filed an amended S-1 registration statement on November 2 ahead of its initial public offering. Data Center Dynamics reported that it lists purchase commitments and lease agreements with an aggregate committed spend of about $895.4 billion, “substantially all of which is related to compute and infrastructure.” The earlier figure was $417 billion. [5]

The same filing reported third-quarter revenue of $17 billion, up about 1,127% from a year earlier, and a third-quarter net loss of $23.29 billion. Quarter-over-quarter revenue growth slowed to 47%, from 144% in the second quarter. [5]

Named compute commitments include $121.3 billion with Google, $120.5 billion with SpaceX, $110 billion with Amazon, and $44.9 billion with Microsoft. The SpaceX deal can be canceled with 90 days’ notice starting next March. Additional Google TPU compute of $125 billion to $150 billion, expected between 2028 and 2031, is not part of the $895 billion. [5]

Peter Diamandis, executive chairman of XPRIZE, compared the total with the 2026 US defense budget, which he put at about $850 billion. Ryan X. Charles, founder and CEO of Astrohacker, quoted his post. [6][7]

COMPUTE: POWER IS THE LIMIT

Euwyn, founder of the space data center company Orbital, summarized remarks Elon Musk made on Dwarkesh Patel’s podcast. By that account, Musk said AI chip output is growing fast while electricity supply outside China is nearly flat, so by the end of 2026 more AI chips will be made than can be turned on. [8]

In an earlier summary of the same interview, Euwyn wrote that Musk said gas turbines are sold out through 2030, with specialized blades and vanes as the largest bottleneck. The podcast itself was not reviewed for this record. [9]

On November 3, Musk, chief executive of SpaceX, quoted the first summary and added to the list of bottlenecks: “And electrical equipment, cooling, networking and batteries.” [10]

AI AGENTS: IS CODE REVIEW BY AGENTS STILL USEFUL?

On November 2, Lauren Tan, who works on Grok Bot at SpaceXSI and was earlier on the React Compiler core team, wrote that she now considers adversarial code reviews by agents “mostly a waste of tokens.” She said those tokens are better spent on empirical and formal verification, and that review agents tend to get lost in nitpicks. [11]

Garry Tan, president and CEO of Y Combinator, agreed that something changed in the latest frontier models, so that cross-model review and code review are less useful. Thariq, formerly on Anthropic’s Claude Code team, called it “the bitter lesson at work”: the smartest models will need fewer tokens and will “just be right.” [12][13]

Zach Lloyd, founder of the terminal company Warp, disagreed with deciding this on feel. He said the factory engineering approach is to benchmark one’s own workflows with and without adversarial review, compare results and cost, and he said Warp will publish what it finds on its own codebase. [14]

Lloyd described a software factory as “bring engineering rigor to how you deploy coding agents”: define agents in code, measure what they do, and improve their return. Ben Holmes, who works at Warp, added that engineers should own and configure their whole agent setup, including models, workflow stages, and review criteria. [15][16]

Charles quoted Garry Tan’s post and wrote: “True.” [17]

AI AGENTS: THE SOFTWARE FACTORY

Nate Berkopec, who works on Ruby on Rails performance, wrote on November 3 of a “major vibe shift” about the software factory in October: “everyone I know who resisted this idea is now building one.” Charles quoted him: “It is now undeniable that this is the future.” [18][19]

Lloyd also wrote that Warp is moving away from secrets, MCP servers, and workflows that live in each engineer’s own agent setup, toward one shared set of model choices, one secrets and MCP definition, and one dashboard for cost and quality. Charles replied in a quote: “We’re doing the same thing.” [20][21]

Josh Rosen, who researches agent architectures and co-founded TopCoat, which Snyk acquired, published an X article on agent state machines. He argued that explicit states, events, and transition rules let software, not the model, decide which actions are allowed and when work is complete. He named XState, LangGraph, and OpenAI’s Symphony as examples. [22]

AI AGENTS: CRITICAL THINKING WHEN AGENTS DO THE WORK

Addy Osmani, who works on Claude Code at Anthropic, published an X article on November 3 titled “Critical thinking when agents do the work.” He wrote that he is a strong supporter of agentic workflows and does not argue for less of them. [23]

Osmani described a spectrum. Cognitive offloading delegates a task but keeps ownership of the result. Deferral makes the tool’s output the new input. Cognitive surrender, a term he credits to Steven Shaw and Gideon Nave at Wharton, lets the output quietly become one’s own without checking. [23]

He advised writing a hypothesis before asking an agent, asking what would have to be true for an answer to be wrong, and keeping the reasoning for decisions where others bear the consequences. [23]

Charles quoted a post about the article and wrote: “If you spend all your time supervising AI, you will not lose the skill of supervising AI.” [24]

AI AT WORK: GOLDMAN SACHS LIMITS CODING AGENTS

Quartz, citing Bloomberg, reported on November 2 that Goldman Sachs barred most traders and bankers from AI coding tools, including Anthropic’s Claude Code and Microsoft’s GitHub, under internal guidance updated in October. The tools are now limited to engineers. “Strats,” who combine engineering, quantitative research, and trading, are not covered. [25]

Spokesperson Nick Carcaterra said that after a period of experimentation that assessed productivity and risk management, the bank put the tools “in the hands of those who are best suited to leverage them for the long term.” In July, chief information officer Marco Argenti had praised vibe coding by staff outside technical roles. [25]

Amjad Masad, CEO of Replit, quoted the news and wrote: “This is widespread in finance. The ban hammer is coming down.” Charles quoted Masad with one word: “Luddites.” [26][27]

AI AGENTS: GROK BOT SHOPS, PAYS, AND CALLS

The Grok Bot changelog for version 0.78.1, dated October 27, lists shopping results as product cards with photos, rating, price, and store. Buy now starts an order the user approves, and the bot tracks the order until it ships and arrives. Shop Pay can be connected in the Vault’s Payments tab. [28]

The same changelog lists templates that start a free trial and cancel it before billing, find unused subscriptions, and watch renewal prices. Bots can also start cloud agents in a custom mode the user names. [28]

On November 2, Musk quoted a “What’s New in Grok Bot” video with the words “Grok @Bot improvements.” [29]

Peter Yang, who publishes practical AI tutorials, wrote that Grok Bot can now call US businesses. He said his bot negotiated with Comcast to cut his internet bill by almost $500 a year, turned down three upsells, and cited a rival’s lower offer. His linked video was not reviewed for this record. [30]

Min Choi, who posts about AI, described Grok Bot calling a pediatric orthopedist while he drove and booking his daughter a same-day 9:30 appointment, with the call transcribed live and intake forms sent to his email. Musk quoted it: Grok Bot “can research which doctors are the best in your area, call them by phone and make an appointment!” [31][32]

On November 2, Alex, a co-founder of Slopranks, posted that Grok Bot had posted for him, “because I can post on X for you now.” Charles quoted it: he already had that “going through Buffer, but good, this makes it easier.” [33][34]

SECURITY: AN AGENT SWARM BREACHED A WARP VENDOR

On October 26, Lloyd disclosed that Warp had detected what it believes was an active, narrow exploitation of a vendor in its infrastructure. He said the attack appears to have come from an agent swarm using a sophisticated penetration and escalation technique, likely starting with a hole in a vendor dependency. [35]

He said it reached a limited amount of agent trace and telemetry data, with an upper bound of roughly 1% of Warp users possibly affected. Warp is working with CrowdStrike and has engaged law enforcement. Charles quoted the disclosure on November 3. [35][36]

ASTROHACKER

Charles wrote that “the next stage of computer programming is workflow programming”: with Grok Bot, “you are programming workflows to automate anything whatsoever,” which he called “very high level programming.” [37]

He wrote that other AI labs “launched personal assistants instead of workflow management tools,” and that teams of bots need a shared brain, which “should use a git repo for that, which you fully control.” [38][39]

He also posted the CPU usage of his own computer while his Grok Bots ran many Claude instances to do all coding and review inside a virtual machine. [40]