- Future Tools
- Posts
- GPT-5.5 remembers (mostly)
GPT-5.5 remembers (mostly)
Plus, Samsung joins the trillion club
Welcome back! Google is updating AI search to pull in "expert advice" from Reddit and other web forums. The idea: surface excerpts from public discussions and perspectives on questions that don't have a single right answer—the kind of thing people already search for by adding "Reddit" to their queries.
Of course, this is the same Google AI Overview feature that once told users to eat rocks and put glue on pizza. Is crowdsourced wisdom the fix, or a new way to hallucinate?
Are We Getting Closer to Truly Proactive AI?
Anthropic just announced something new this week called Dreaming. It's part of the Managed Agents platform, so it's not something you’ll see in the normal Claude app right away.
This isn't exactly like the memory we already have in something like ChatGPT or Claude, where it remembers past conversations, saves little details it thinks are important, and then brings some of that back into context the next time you chat.
This seems different:
It happens on a schedule, reviews past memories and past conversations, looks for patterns, spots mistakes that keep popping up, and notices things you're doing over and over again.
Then, ideally, it can proactively figure out ways to help solve those problems for you.
My POV: This is where we all want AI assistants to eventually end up.
We don't want to always have to think of a problem, open a chat, explain the issue, and ask the AI to solve it. We want these assistants to learn our businesses, habits, and mistakes, then proactively suggest ways to help.
“Dreaming” feels like a step in that direction. It's getting us closer to an AI assistant that comes to us and says, "Here's what I've noticed. Let me help you fix it," instead of always waiting for us to ask.
It's an exciting development, and I really wish it was already available inside the normal Claude experience…but for now it's only part of Managed Agents.
— Matt
GPT-5.5 Instant Shows You (Some of) What It Remembered
Via OpenAI
OpenAI updated ChatGPT's default model to GPT-5.5 Instant—and with it, debuted a new feature that shows users which memories and past chats shaped a response. But there's a catch: it doesn't show everything.
What's new: When ChatGPT personalizes a response, you can now tap a "sources" button to see which of your saved memories or past conversations the model drew from. Delete or correct outdated context, and if you forward a conversation, your sources won't be shared to others.
The limitation: OpenAI admitted the feature "may not show every factor that shaped an answer." For enterprises, this creates a potential headache: model-reported context could overlap or contradict existing retrieval logs, making it harder to trace failures or audit responses reliably.
The model itself: Internal evaluations showed GPT-5.5 Instant returned 52.5% fewer hallucinations than the previous default, with inaccurate claims down 37.3% on challenging conversations, especially in high-stakes domains like medicine, law, and finance.
Why it matters: AI models have gotten very personal—remembering your preferences, past questions, and files. But how much memory is helpful, and when does it just dredge up irrelevant context from months ago? Transparency is a start, but partial observability isn't the same as auditability.
AI Optimism Pushes Chip Stocks to Record Highs
AI demand is fueling a surge across the semiconductor industry, with two major milestones this week: Samsung hit a $1 trillion valuation, and AMD shares jumped to an all-time high.
Samsung crosses $1 trillion: Shares of the South Korean giant jumped more than 10% on Wednesday, making it only the second Asian company to hit the trillion-dollar mark after TSMC. The catalyst: profits eight times higher than a year ago, driven by soaring demand for high-bandwidth memory chips that power AI systems. Reports that Apple is in talks with Samsung to manufacture chips on US soil also added to the massive valuation.
AMD hits record highs: AMD shares surged 14.9%, adding more than $86 billion in market value in a single session. The company raised its forecast for the server CPU market, now expecting 35% annual growth through 2030—up from 18% previously. Analysts see AMD as the leading challenger to Nvidia, and demand keeps growing as companies shift toward agentic AI.
The ripple effect: Intel, Arm, Qualcomm, ASML, and Super Micro all rallied. The Philadelphia SE Semiconductor index hit a fresh record—up 59% year-to-date, including a 38% jump in April alone, its best month since 2000.
The bigger picture: Chipmakers are pulling investment from consumer products to ramp up AI infrastructure, and the math is working: S&P 500 semiconductor companies are expected to post 109.8% earnings growth for Q1, the highest of any tech sub-sector.
Give your AI agents eyes and ears
Via VideoDB
VideoDB ingests live or recorded audio and video, automatically transcribes and extracts visual context, and streams low-latency data to AI agents. It builds multimodal vector indices so agents can search, recall moments, trigger alerts, and act on media programmatically.
How you can use it
Power meeting copilots, screen-aware coding assistants, or monitoring workflows
Semantically search video libraries by content, not just metadata
Offload heavy media processing to a dedicated infrastructure layer
Integrate with any LLM via SDKs and APIs
Pricing: Free and paid plans
AI-powered document archive
Via Veluvanto
Veluvanto automatically reads, extracts, tags, and organizes invoices, contracts, receipts, and other files. Ask questions in natural language to get verifiable answers with source documents attached.
How you can use it
Stop hunting for paperwork—search your archive instantly
Auto-extract key fields like amounts, dates, and parties
Detect deadlines and set up approval workflows
Get charts and summaries with full audit trails
Pricing: Free and paid plans
Jobs, announcements, and big ideas
OpenAI ships GPT-5.5-Cyber Preview, a trusted-access model built for critical infrastructure defenders.
Anthropic donates Petri, its open-source AI alignment tool, to Meridian Labs for continued development.
Perplexity launches its Personal Computer agent for all Mac users in a new macOS app.
Apple nears mass production of AirPods with built-in AI cameras.
OpenAI embeds Codex directly inside Chrome on macOS and Windows for in-browser coding agents.
Google releases Gemini 3.1 Flash-Lite to general availability across Vertex AI and the Gemini API.
Spotify debuts a beta tool that lets users save and revisit their AI-generated personal podcasts.
Want a rundown on the week’s biggest tech headlines? I’ve got you covered.
That’s a wrap! See you next week for more.