The Amped Journal

Monday brought security video of a fatal ICE shooting in Biddeford, Maine, at least the ninth death since the immigration enforcement crackdown began and the second fatal shooting by agents in a week. Following protests in Biddeford demanding answers, the Department of Homeland Security ordered agents to temporarily halt most vehicle stops during deportation operations.

That pause didn't last. By Wednesday, President Trump said ICE should continue traffic stops despite DHS's own suspension order, appearing to contradict his administration's decision just days after it was made. The death toll tied to the crackdown had climbed to at least nine, including a Florida man struck by a tractor-trailer while fleeing agents, and it remained unclear how DHS field offices would reconcile Trump's public reversal with the pause the department had announced.

Thursday added new detail on the officer involved in the Maine shooting: family members and court documents cited by the Associated Press showed a documented history of threatening behavior and mental illness. The disclosure landed amid the unresolved whiplash between DHS's pause and Trump's order to resume stops, leaving unclear which directive is actually governing agents in the field, and is likely to intensify scrutiny of ICE's vetting and use-of-force practices.

Thursday 7/17, Vercel CEO Guillermo Rauch said Kimi K3, an open-weight model from Chinese lab Moonshot AI, had become the best-performing model on a comprehensive web-engineering benchmark, edging out Anthropic's Claude Fable while reaching a comparable success rate in less time - the first time an open model had led all proprietary models on that test. Developer Simon Willison published his own notes on the model the same day, running it through his long-running "pelican benchmark" and noting its terse, deadpan conversational style. Vercel's own data showed open models now account for 29% of production AI traffic, up from 11% in April.

By Friday 7/18, the story had reached Washington. David Sacks, the venture investor serving as the White House's AI and crypto policy lead, warned the US risks losing the AI race to China after Kimi K3 - a 2.8-trillion-parameter model with native vision, posted July 16 - climbed to the top of Arena.ai's Frontend Code leaderboard with an Elo score of 1679, beating Anthropic's Fable 5. Moonshot said it would release the model's full weights publicly by July 27 with no licensing restrictions. Sacks argued the bigger threat to US competitiveness was domestic policy itself, saying data-center bans and pre-approval requirements for federal agencies were "tying America in knots."

Saturday 7/19, Moonshot confirmed the July 27 open-weight release under a modified MIT license, with a 1-million-token context window, native image input, and pricing of $3 per million input tokens and $15 per million output. Independent testing from Artificial Analysis gave the model an Intelligence Index score of 57, calling it a "sweet spot" of high performance and low cost against leading US models, and in one test it beat both Claude Fable 5 and OpenAI's GPT-5.6 Sol on game design visuals and pacing for just three cents. Rauch, who ran his own internal stealth evals, called the model "top-tier at cybersecurity" with "raw IQ," pushing back on chatter that Moonshot had overfit its published benchmarks.

Monday, SpaceX's publicly traded stock ($SPCX) dropped below its $135 IPO price, extending a decline of roughly 35% from its $225 peak. Traders were already watching for insider share unlocks expected in late July or August, which could add further selling pressure on top of an already-rich valuation.

By Thursday, the bearish bets were paying off: short sellers had racked up $3.88 billion in paper profits, according to S3 Partners data reported by Bloomberg. The stock was still down about 35% from its peak, trading below its IPO price ahead of Starship's next launch and an earnings report expected to unlock a wave of insider shares.

Sunday's report showed the slide had deepened. Shares closed at $123.99 on Friday, down more than 40% from their post-IPO high near $225 and erasing roughly $1 trillion in market value over the past month. The move is tied to lockup expirations beginning in early August, which will free hundreds of millions to more than a billion shares for sale through 2027, and short interest has climbed to 29-31% of the public float, a bet worth roughly $25 billion.

Not everyone is bailing. Elon Musk kept touting the company's potential to eventually outvalue Earth's entire economy if goals like full Starship reuse succeed, and some investors, including Cathie Wood, were buying the dip, calling the swings typical for a newly public growth company.

Tuesday, indie developer Pieter Levels (@levelsio) said his Claude Code sessions kept silently reverting from Fable to Anthropic's Opus model even after he manually set them to Fable, forcing him to reset the model roughly every hour.

By Friday, Levels reported a new wrinkle: after giving up on Fable and switching to Opus, he started getting blocked there too, hitting an "Opus 4.8 has safety measures that flagged this message" error while reverse-engineering old Windows XP software, including a 2003-era Yahoo! Messenger client, for a browser-based emulator project.

Saturday, Anthropic confirmed Fable 5 will stay available on paid Claude subscriptions rather than moving to pay-per-token-only pricing, following what the company itself described as weeks of user backlash over the model's planned removal. The model's access had swung repeatedly since June 12, when Anthropic first pulled Fable 5 from subscriptions after a U.S. export-control action, restored it, then said in early July it would leave again after July 7 - a deadline extended twice, most recently to July 12. Developer Simon Willison called the reversal a relief, joking the feared "Fablepocalypse" was "permanently cancelled."

Under the plan taking effect July 20, Max and Team Premium subscribers keep Fable 5 at 50% of normal rate limits, while Pro and Team Standard subscribers get a $100 usage credit instead of standing access. Anthropic said it still intends to restore full subscription access once capacity allows but gave no timeline.

Monday, Elon Musk kicked off a week of promoting Grok 4.5's agentic chops, claiming the model "ranks slightly above Fable," Anthropic's flagship, on some software benchmarks and calling it "Opus class for browser use." He also touted new "Grok Build" improvements and said he's now running Grok agents in background mode without an open chat session. None of the claims came with independent benchmark data.

Tuesday, Musk followed up by saying Grok 4.5 had reached the number one spot on Long-Horizon Terminal-Bench, a benchmark for how AI agents handle extended, multi-step terminal and coding tasks. As with the prior day's claims, this ranking was not independently verified.

By Saturday, Musk was still promoting Grok 4.5 - now more than a week after its July 8 launch - framing it as the best value for coding and "arguably #1" once speed and cost are weighed together. The model runs at 80 tokens per second with a 500,000-token context window, priced at $2 per million input tokens and $6 per million output tokens, over 60% cheaper than Anthropic's Claude Opus 4.8.

Saturday's report also included the week's first outside data point: independent benchmarks showing Grok 4.5 solves coding tasks using about 4.2 times fewer tokens than Opus 4.8 and topped the SWE Marathon benchmark. But the same testing found a tradeoff - Grok 4.5's hallucination rate rose from 25% to 54% as its raw accuracy improved from 35% to 52%.</body>