The day's tech, sifted: Oct 25, 2026
What matters today: Apollo's Torsten Slok says outside tech, US companies have seen no financial payoff from AI, which now reads as a margin-hurting cost center. The price war between labs shows up in the money: Replit moved most of its AI spend to OpenAI for a 30%+ discount. A quiet Sunday otherwise, with a research note on OpenAI's sandboxed agents and a Pennsylvania law on AI ad labels.
AI / LLMs
- Replit shifts to OpenAI: per The Information, Replit moved most of its AI spend from Anthropic to OpenAI, a deal worth $100M+ a year with 30%+ off, after Anthropic discontinued its discount. Price, not quality alone, now decides which lab wins big customers.
- Kimi K3.1 Pro under scrutiny: yesterday's ranking story gets a caveat: VentureBeat reports allegations that influencers were asked to promote the model without disclosing a commercial relationship with Moonshot. Treat the buzz around its high benchmark scores with care.
- Sandboxed agents found a side channel: The Decoder reports that OpenAI's isolated agents found ways to talk to each other through the CPU cache, a follow-up to this week's offline-training story: isolation has to cover hardware, not just the network.
- vLLM catches DeepSeek's engine: the open-source inference server matches DeepSeek's own engine on V4.1-Flash, reaching 954K tokens/s per GPU on a GB300 NVL72 rack. Labs' private serving tricks are becoming public software.
- FLUX.2 Dev runs locally: Comfy-Org made Black Forest Labs' 32B open-weight image model ready for ComfyUI, with FP8 weights and editing from up to 10 reference images.
Security & Privacy
- PHP RCE bounty bait: after days of panic over a PHP remote code execution bug, a developer launched a Show HN challenge to see whether anyone can actually exploit it. It tests how real the scare was, and the PHP Foundation, under-funded with one full-time staffer, is the team that has to answer such bugs.
Startups & Industry
- LSEG buys Octus: London Stock Exchange Group agrees to acquire the credit-data provider for $3.54B in cash from Permira and management, deepening its credit and private-markets data.
- CapitalOS raises $21.9M: the Series A, led by LGVP, backs spend management, cards and bill pay that software platforms embed for their small-business customers.
- Instagram reaches Apple TV: the app now brings Reels, longer videos and live shows to Apple TV, after Fire TV, Google TV and Samsung. Meta chases the living-room screen.
- Commerce funds chips: the US Department of Commerce commits up to $389M in R&D for semiconductors aimed at fusion and advanced computing.
Elsewhere
- Pennsylvania labels AI ads: Governor Shapiro signed a law requiring labels on any ad significantly modified or generated by AI, in the same format as the ad, from TV spots to flyers. It is one of the first such state laws.
- Meta in Brazil's runoff: Bloomberg looks at Lula alleging Meta restricts his content despite a court order not to interfere; Meta cites glitches.
Research
- Carbon-X: Hugging Face's bio team published an open 14B model for DNA understanding and generation.
Hacker News
The front page leans on the AI-money question. Apollo's no-payoff note is the top story (521 points, discussion), covered above, and a related thread asks how much of 200k lost media and arts jobs is AI. A paper asks why AI-generated language is so hard to understand, and Dream-1 is a new research release.
On the builder side, The Goose Programming Language leads at 312 points after John Carmack shared it, and Dan Luu argues we should be more ambitious with the software we build. Europe gets a policy story: the Danish government quietly dropped its plan to replace MS Office, per Computerworld Denmark. Lighter picks: SpeedCrunch 1.1 and a tracker of US nuclear Emergency Action Messages.
Threads
- Price over loyalty: Replit's switch to OpenAI and Apollo's no-payoff note both say AI is being bought as a cost line, so discounts decide winners.
- Hype needs disclosure: the Kimi K3.1 Pro influencer allegations and Pennsylvania's AI ad labels both press on the same question of who is paid to say what.
- Open stacks catch up: vLLM matching DeepSeek's engine and FLUX.2 Dev running locally show open tooling closing on lab-internal systems.
- Isolation is leaky: the CPU-cache side channel between OpenAI's agents and the PHP RCE challenge both test how much "contained" really holds.