prompt 🤖 AI News
13.5K subscribers
57 photos
24 videos
1 file
939 links
Welcome to @prompt, your go-to source for AI insights, breakthroughs, and tools shaping the future of intelligence.


Contact: @LightEarendil
Download Telegram
🤖 Lovable deleted most of its agent code, and says it'll keep doing that

After 100k PRs, the team wrote up why they keep rebuilding their coding agent. Each new base model makes chunks of their scaffolding redundant, so the harness works like a depreciating asset. Prompt tricks and orchestration glue go stale in months.

Whatever you hand-built last quarter is next quarter's cleanup ticket.
🧠 HazyResearch wants LLMs to generate concepts instead of tokens, in parallel

Next-Concept Modeling generates text coarse-to-fine, so it drafts the high-level idea first and fills in detail after, rather than decoding one token at a time. The repo is public, but I haven't seen independent speed or quality numbers against autoregressive baselines yet.
🚨 Nvidia opens an investigation into Jensen Huang over ties to a Chinese spy

The NYT reports that Huang told the board he'd been intimate with a business executive. U.S. intelligence determined she was a Chinese spy.

Nvidia announced the probe Sunday. Huang is the CEO and co-founder of the company that supplies most of the chips used to train frontier AI.
😘2
🏛️ The US AI safety center's staff are now reporting to the White House

AP reports, citing two former officials, that CAISI staff were recently asked to informally report to a White House office, even though the center is formally an independent Commerce Department division with a budget of about $10M a year, roughly one-eighth of the UK's, and Congress asked in August why it stopped publishing its research.
🚨 Voters shot down data center ballot measures in 40 of 41 cases, right as Anthropic heads to IPO

The midterm results, per CNBC, land at an awkward moment for a lab that needs a lot of new buildings. Anthropic's own prospectus lists organized opposition to data centers as a risk factor, and says its public profile could draw extra scrutiny to projects tied to the company.
🇪🇸 A Spanish town is fighting a macro data center, and residents call it a way of kicking them out

El País reports that the pushback has moved from US suburbs to rural Europe.

• The fight is over land, water and who benefits
• The town is small, the developer isn't
• Locals say the project would push them out
❤1
⚡️ Google's new TPUs 8tx and 8i Max are built for agents, not chat

Google announced the two chips with a software layer called Agent Compute Engine. It suspends idle agent sessions and moves working memory off the chip in milliseconds, so waiting on a database or a human stops burning power.

No availability or pricing yet. And Google hasn't said how the off-chip move stays fast enough for pauses of a few seconds.
🚨 The EU wants to regulate addictive design in AI chatbots, not just social media

Bloomberg reports that a draft law targets engagement hooks in both feeds and chatbot products. The focus is shifting from what models say to how products keep people hooked.

It's still a draft, so the final text could change. Companion-style apps and big assistants would be the obvious ones to feel it.
🤖 Meta's Muse Ember is built to be the cheap worker bee for agents

Meta Superintelligence Lab released its lightest, most cost-efficient model yet, meant to pair with a bigger one like Muse Wonder.

• Target jobs: tool calls, coding subtasks, routing, extraction, long document parsing
• Reads screenshots, scanned docs, voice memos and video without a separate model
• Fourth Muse model, after Spark (April), Glimmer (August) and Wonder (October)
🤖 Anthropic ships Fable 5.2 and Mythos 5.2 together, one version bump for both lines

Anthropic posted the update just weeks after the 5.0 launch, where Fable was the general release and Mythos was the cyber-defense tier for vetted users, with the same underlying model and different safeguards. I couldn't verify 5.2 benchmarks or pricing yet, so whether that split still holds is the open question.
🚨 Fireworks AI says someone used its internal credentials to reach customer accounts

Fireworks disclosed that the attacker read environment variables used in evaluation jobs, in a limited number of accounts. It spotted the activity Oct 6 and has since revoked affected API keys and rotated employee and service-account tokens.

It says it found no sign of access to customer models, datasets, training data or inference traffic.
❤1
🚨 Fireworks AI gets hacked again, second breach in weeks

Bloomberg reports the Nvidia-backed inference provider has notified customers of a new incident.

• The October breach exposed environment variables in eval jobs
• Fireworks serves other companies' models, which means it holds their keys
• Rotating credentials is the obvious move for anyone on the platform
🚨 Claude Fable 5.2 takes #1 on Artificial Analysis with a 65, and the bill doubled

Artificial Analysis ranks it the highest-scoring model so far, but that's at Max effort.

• Output tokens per task jump from 118k to 274k vs Fable 5.1
• Reasoning tokens alone go from 71k to 192k
• At High effort it uses about 80k tokens
• It beats Opus 5.5 Max at lower cost
🚨 Anthropic wants $69.3B from its IPO, which would be one of the biggest tech listings ever

Reuters reports the raise target, and it lands as a lab stops being a private bet.

• Public filings mean real numbers on compute spend and margins
• Quarterly earnings pressure meets a safety-first brand
• OpenAI's last round valued it at $1.25T, so that's the bar
🚨 400+ "normal" AI apps hide explicit deepfake tools behind a server switch

Wired traced a connected network of apps and domains that look harmless in review, then unlock the abusive stuff later.

• Disguises include pet-photo generators, language classes and AI workout coaches
• A server check flips on the hidden features
• One trigger was installing the app after tapping an ad on Meta platforms
🚨 CoreWeave's 2032 bond now yields over 16%

CNBC reports the 9.625% coupon paper is trading well under par, and five-year CDS hit 1,289 bps Thursday, above the October peak. Protection is also getting pricier on Nvidia, Broadcom and Oracle. Google's bonds yield around 5.7%.

Jefferies' Chris Wood says the AI capex race ends when the credit card gets taken away.
🔐 DeepMind's AI agents just formally verified Signal's actual protocol code

DeepMind says AlphaProof Code, with Gemini Carbon agents doing the proofs, wrote over six million lines of Lean. Humans reviewed every spec. It's the first formal verification of a widely deployed protocol's implementation, with or without AI.

The agents proved the code matches the spec, but people still decided what "secure" means.
🚨 Anthropic's IPO is pricing at $1.1T to $1.3T, well short of the $2T some backers wanted

BBC reports the range is $70 to $78 a share, raising up to $64bn on Nasdaq. Revenue hit $16.2bn for the first half of the year, up from $1.1bn a year earlier.

It still lands below SpaceX, which listed in June at nearly $1.8T.
🚨 Gemini 4 Carbon is live, and Google says it's the strongest Gemini yet

Google says it ranks #1 on CursorBench 4.0, WANDR and AutomationBench at more than 2X lower cost per task than the next best external model. It burns more tokens than Argon to reason deeper.

The API is open, but in the Gemini app it looks Ultra-only. Pro subscribers are still staring at 3.1 Pro.
🖼️ MiniMax opens the weights on H3-Image, a 33B model that renders 16 megapixels without an upscaler

MiniMax released one checkpoint for text-to-image and editing, distilled to 6 sampling steps with no CFG, and it takes up to 16 reference images. The DiT is 68.65 GB in BF16, and it reuses the H3 text encoder, which is a separate 66.7 GB download.
🚨 Anthropic is going after Microsoft Office, and it's doing it right before the IPO

Bloomberg reports the launch of Claude Whiteboards, Claude Sheets and Claude Data, which move brainstorming, spreadsheets and shared data into Claude itself, so enterprise files live there instead of in a chat window next to Copilot.
❤1